@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
|
@@ -1,5 +1,5 @@
|
|
|
1
1
|
import { buildRuntimeContractPatch } from './command-api.js';
|
|
2
|
-
import {
|
|
2
|
+
import { applyOrderedRuntimeInputs } from './runtime-input-composition.js';
|
|
3
3
|
|
|
4
4
|
function cloneRuntimeConfig(runtimeConfig) {
|
|
5
5
|
if (runtimeConfig == null) return runtimeConfig;
|
|
@@ -9,6 +9,46 @@ function cloneRuntimeConfig(runtimeConfig) {
|
|
|
9
9
|
return JSON.parse(JSON.stringify(runtimeConfig));
|
|
10
10
|
}
|
|
11
11
|
|
|
12
|
+
function assertCalibrateRuntimeCompatibility(request, runtimeConfig) {
|
|
13
|
+
if (request?.intent !== 'calibrate') {
|
|
14
|
+
return;
|
|
15
|
+
}
|
|
16
|
+
|
|
17
|
+
const shared = runtimeConfig?.shared ?? {};
|
|
18
|
+
const debug = shared.debug ?? {};
|
|
19
|
+
const benchmarkRun = shared.benchmark?.run ?? {};
|
|
20
|
+
const violations = [];
|
|
21
|
+
|
|
22
|
+
if (debug.trace?.enabled === true) {
|
|
23
|
+
violations.push('runtime.shared.debug.trace.enabled');
|
|
24
|
+
}
|
|
25
|
+
if (debug.pipeline?.enabled === true) {
|
|
26
|
+
violations.push('runtime.shared.debug.pipeline.enabled');
|
|
27
|
+
}
|
|
28
|
+
if (Array.isArray(debug.probes) && debug.probes.length > 0) {
|
|
29
|
+
violations.push('runtime.shared.debug.probes');
|
|
30
|
+
}
|
|
31
|
+
if (debug.profiler?.enabled === true) {
|
|
32
|
+
violations.push('runtime.shared.debug.profiler.enabled');
|
|
33
|
+
}
|
|
34
|
+
if (benchmarkRun.debug === true) {
|
|
35
|
+
violations.push('runtime.shared.benchmark.run.debug');
|
|
36
|
+
}
|
|
37
|
+
if (benchmarkRun.profile === true) {
|
|
38
|
+
violations.push('runtime.shared.benchmark.run.profile');
|
|
39
|
+
}
|
|
40
|
+
if (benchmarkRun.captureMemoryTimeSeries === true) {
|
|
41
|
+
violations.push('runtime.shared.benchmark.run.captureMemoryTimeSeries');
|
|
42
|
+
}
|
|
43
|
+
|
|
44
|
+
if (violations.length > 0) {
|
|
45
|
+
throw new Error(
|
|
46
|
+
`tooling command: calibrate intent forbids investigation instrumentation (${violations.join(', ')}). ` +
|
|
47
|
+
'Disable those runtime config fields or use the debug command instead.'
|
|
48
|
+
);
|
|
49
|
+
}
|
|
50
|
+
}
|
|
51
|
+
|
|
12
52
|
function resetRuntimeState(runtimeBridge) {
|
|
13
53
|
if (!runtimeBridge?.setRuntimeConfig) {
|
|
14
54
|
throw new Error('runtime bridge must provide setRuntimeConfig().');
|
|
@@ -22,12 +62,6 @@ function resetRuntimeState(runtimeBridge) {
|
|
|
22
62
|
runtimeBridge.setRuntimeConfig(null);
|
|
23
63
|
}
|
|
24
64
|
|
|
25
|
-
function mergeRuntimePatch(runtimeBridge, patch) {
|
|
26
|
-
if (!patch) return;
|
|
27
|
-
const mergedRuntime = mergeRuntimeValues(runtimeBridge.getRuntimeConfig(), patch);
|
|
28
|
-
runtimeBridge.setRuntimeConfig(mergedRuntime);
|
|
29
|
-
}
|
|
30
|
-
|
|
31
65
|
function snapshotRuntimeState(runtimeBridge) {
|
|
32
66
|
return {
|
|
33
67
|
runtimeConfig: cloneRuntimeConfig(runtimeBridge.getRuntimeConfig()),
|
|
@@ -73,17 +107,17 @@ function restoreRuntimeState(runtimeBridge, snapshot) {
|
|
|
73
107
|
|
|
74
108
|
export async function applyRuntimeInputs(request, runtimeBridge, options = {}) {
|
|
75
109
|
resetRuntimeState(runtimeBridge);
|
|
76
|
-
|
|
77
|
-
|
|
78
|
-
|
|
79
|
-
|
|
80
|
-
|
|
81
|
-
|
|
82
|
-
|
|
83
|
-
|
|
84
|
-
|
|
85
|
-
|
|
86
|
-
|
|
110
|
+
await applyOrderedRuntimeInputs(runtimeBridge, {
|
|
111
|
+
configChain: request.configChain ?? null,
|
|
112
|
+
runtimePreset: request.runtimePreset ?? null,
|
|
113
|
+
runtimeConfigUrl: request.runtimeConfigUrl ?? null,
|
|
114
|
+
runtimeConfig: request.runtimeConfig ?? null,
|
|
115
|
+
runtimeContractPatch: buildRuntimeContractPatch(request),
|
|
116
|
+
}, {
|
|
117
|
+
applyRuntimePreset: runtimeBridge.applyRuntimePreset?.bind(runtimeBridge),
|
|
118
|
+
applyRuntimeConfigFromUrl: runtimeBridge.applyRuntimeConfigFromUrl?.bind(runtimeBridge),
|
|
119
|
+
}, options);
|
|
120
|
+
assertCalibrateRuntimeCompatibility(request, runtimeBridge.getRuntimeConfig());
|
|
87
121
|
}
|
|
88
122
|
|
|
89
123
|
export async function runWithRuntimeIsolation(runtimeBridge, run) {
|
|
@@ -1,12 +1,159 @@
|
|
|
1
1
|
import {
|
|
2
|
-
extractExecutionContractFacts,
|
|
3
2
|
sanitizeLeanModuleName,
|
|
4
3
|
} from '../config/execution-contract-check.js';
|
|
4
|
+
import { DEFAULT_BATCHING_DEFAULTS, DEFAULT_GENERATION_CONFIG } from '../config/schema/inference-defaults.schema.js';
|
|
5
|
+
import { DEFAULT_KVCACHE_CONFIG } from '../config/schema/kvcache.schema.js';
|
|
5
6
|
|
|
6
7
|
export {
|
|
7
|
-
extractExecutionContractFacts,
|
|
8
8
|
sanitizeLeanModuleName,
|
|
9
|
-
}
|
|
9
|
+
};
|
|
10
|
+
|
|
11
|
+
const KV_LAYOUTS = new Set(['contiguous', 'paged', 'tiered', 'bdpa', 'bdpa_paged']);
|
|
12
|
+
const PHASES = new Set(['prefill', 'decode', 'both']);
|
|
13
|
+
const COLD_QUANT_MODES = new Set(['none', 'int8', 'int4']);
|
|
14
|
+
const ATTENTION_OPS = new Set(['attention']);
|
|
15
|
+
const EMBED_OPS = new Set(['embed', 'gather']);
|
|
16
|
+
const SAMPLE_OPS = new Set(['sample']);
|
|
17
|
+
|
|
18
|
+
function isPlainObject(value) {
|
|
19
|
+
return value != null && typeof value === 'object' && !Array.isArray(value);
|
|
20
|
+
}
|
|
21
|
+
|
|
22
|
+
function assertManifestObject(manifest) {
|
|
23
|
+
if (!isPlainObject(manifest)) {
|
|
24
|
+
throw new Error('lean execution contract: manifest must be an object.');
|
|
25
|
+
}
|
|
26
|
+
return manifest;
|
|
27
|
+
}
|
|
28
|
+
|
|
29
|
+
function assertPositiveInteger(value, label) {
|
|
30
|
+
if (!Number.isInteger(value) || value < 0) {
|
|
31
|
+
throw new Error(`lean execution contract: ${label} must be a non-negative integer.`);
|
|
32
|
+
}
|
|
33
|
+
return value;
|
|
34
|
+
}
|
|
35
|
+
|
|
36
|
+
function assertPositiveIntegerOrDefault(value, fallback, label) {
|
|
37
|
+
if (value == null) {
|
|
38
|
+
return assertPositiveInteger(fallback, `${label} fallback`);
|
|
39
|
+
}
|
|
40
|
+
return assertPositiveInteger(value, label);
|
|
41
|
+
}
|
|
42
|
+
|
|
43
|
+
function normalizeKVLayout(value) {
|
|
44
|
+
const normalized = String(value ?? DEFAULT_KVCACHE_CONFIG.layout).trim().toLowerCase();
|
|
45
|
+
if (!KV_LAYOUTS.has(normalized)) {
|
|
46
|
+
throw new Error(
|
|
47
|
+
`lean execution contract: unsupported KV layout "${value}". ` +
|
|
48
|
+
`Expected one of ${[...KV_LAYOUTS].join(', ')}.`
|
|
49
|
+
);
|
|
50
|
+
}
|
|
51
|
+
return normalized;
|
|
52
|
+
}
|
|
53
|
+
|
|
54
|
+
function normalizePhase(value, label) {
|
|
55
|
+
const normalized = String(value ?? '').trim().toLowerCase();
|
|
56
|
+
if (!PHASES.has(normalized)) {
|
|
57
|
+
throw new Error(
|
|
58
|
+
`lean execution contract: ${label} must be one of ${[...PHASES].join(', ')}.`
|
|
59
|
+
);
|
|
60
|
+
}
|
|
61
|
+
return normalized;
|
|
62
|
+
}
|
|
63
|
+
|
|
64
|
+
function normalizeColdQuantMode(kvcache) {
|
|
65
|
+
const tieringMode = String(
|
|
66
|
+
kvcache?.tiering?.mode
|
|
67
|
+
?? DEFAULT_KVCACHE_CONFIG.tiering.mode
|
|
68
|
+
).trim().toLowerCase();
|
|
69
|
+
if (tieringMode === 'off' || tieringMode === 'fp16') {
|
|
70
|
+
return 'none';
|
|
71
|
+
}
|
|
72
|
+
if (!COLD_QUANT_MODES.has(tieringMode)) {
|
|
73
|
+
throw new Error(
|
|
74
|
+
`lean execution contract: unsupported tiered cold quant mode "${tieringMode}".`
|
|
75
|
+
);
|
|
76
|
+
}
|
|
77
|
+
return tieringMode;
|
|
78
|
+
}
|
|
79
|
+
|
|
80
|
+
function classifyOp(op) {
|
|
81
|
+
const normalized = String(op ?? '').trim().toLowerCase();
|
|
82
|
+
if (!normalized) {
|
|
83
|
+
throw new Error('lean execution contract: execution step op is required.');
|
|
84
|
+
}
|
|
85
|
+
if (ATTENTION_OPS.has(normalized)) return 'attention';
|
|
86
|
+
if (EMBED_OPS.has(normalized)) return 'embed';
|
|
87
|
+
if (SAMPLE_OPS.has(normalized)) return 'sample';
|
|
88
|
+
if (normalized.includes('norm')) return 'norm';
|
|
89
|
+
if (normalized.includes('residual')) return 'residual';
|
|
90
|
+
if (normalized.endsWith('_proj') || normalized.startsWith('rope_') || normalized === 'activation') {
|
|
91
|
+
return 'projection';
|
|
92
|
+
}
|
|
93
|
+
return 'other';
|
|
94
|
+
}
|
|
95
|
+
|
|
96
|
+
export function extractExecutionContractFacts(manifest) {
|
|
97
|
+
const normalizedManifest = assertManifestObject(manifest);
|
|
98
|
+
const modelId = String(normalizedManifest.modelId ?? 'model').trim() || 'model';
|
|
99
|
+
const architecture = isPlainObject(normalizedManifest.architecture)
|
|
100
|
+
? normalizedManifest.architecture
|
|
101
|
+
: {};
|
|
102
|
+
const inference = isPlainObject(normalizedManifest.inference)
|
|
103
|
+
? normalizedManifest.inference
|
|
104
|
+
: {};
|
|
105
|
+
const sessionDefaults = isPlainObject(inference.sessionDefaults)
|
|
106
|
+
? inference.sessionDefaults
|
|
107
|
+
: {};
|
|
108
|
+
const execution = isPlainObject(inference.execution)
|
|
109
|
+
? inference.execution
|
|
110
|
+
: {};
|
|
111
|
+
const kvcache = isPlainObject(sessionDefaults.kvcache)
|
|
112
|
+
? sessionDefaults.kvcache
|
|
113
|
+
: {};
|
|
114
|
+
const decodeLoop = isPlainObject(sessionDefaults.decodeLoop)
|
|
115
|
+
? sessionDefaults.decodeLoop
|
|
116
|
+
: {};
|
|
117
|
+
|
|
118
|
+
const steps = Array.isArray(execution.steps)
|
|
119
|
+
? execution.steps.map((step, index) => {
|
|
120
|
+
if (!isPlainObject(step)) {
|
|
121
|
+
throw new Error(`lean execution contract: execution.steps[${index}] must be an object.`);
|
|
122
|
+
}
|
|
123
|
+
const id = String(step.id ?? '').trim();
|
|
124
|
+
if (!id) {
|
|
125
|
+
throw new Error(`lean execution contract: execution.steps[${index}].id is required.`);
|
|
126
|
+
}
|
|
127
|
+
return {
|
|
128
|
+
id,
|
|
129
|
+
phase: normalizePhase(step.phase, `execution.steps[${index}].phase`),
|
|
130
|
+
opClass: classifyOp(step.op),
|
|
131
|
+
};
|
|
132
|
+
})
|
|
133
|
+
: [];
|
|
134
|
+
|
|
135
|
+
return {
|
|
136
|
+
modelId,
|
|
137
|
+
session: {
|
|
138
|
+
layout: normalizeKVLayout(kvcache.layout),
|
|
139
|
+
disableCommandBatching: decodeLoop.disableCommandBatching
|
|
140
|
+
?? DEFAULT_GENERATION_CONFIG.disableCommandBatching,
|
|
141
|
+
decodeBatchSize: assertPositiveIntegerOrDefault(
|
|
142
|
+
decodeLoop.batchSize,
|
|
143
|
+
DEFAULT_BATCHING_DEFAULTS.batchSize,
|
|
144
|
+
'sessionDefaults.decodeLoop.batchSize'
|
|
145
|
+
),
|
|
146
|
+
headDim: assertPositiveInteger(architecture.headDim, 'architecture.headDim'),
|
|
147
|
+
kvLen: assertPositiveIntegerOrDefault(
|
|
148
|
+
architecture.maxSeqLen ?? kvcache.maxSeqLen,
|
|
149
|
+
DEFAULT_KVCACHE_CONFIG.maxSeqLen,
|
|
150
|
+
'architecture.maxSeqLen'
|
|
151
|
+
),
|
|
152
|
+
coldQuantMode: normalizeColdQuantMode(kvcache),
|
|
153
|
+
},
|
|
154
|
+
steps,
|
|
155
|
+
};
|
|
156
|
+
}
|
|
10
157
|
|
|
11
158
|
function asLeanString(value) {
|
|
12
159
|
return JSON.stringify(String(value));
|