doppler-gpu 0.4.7 → 0.4.8
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 +64 -0
- package/README.md +93 -144
- package/models/catalog.json +1575 -0
- package/package.json +37 -12
- package/src/cli/doppler-cli.js +28 -7
- package/src/client/doppler-registry.json +2 -2
- package/src/client/runtime/index.js +25 -2
- package/src/client/runtime/model-session.d.ts +31 -1
- package/src/client/runtime/model-session.js +31 -0
- package/src/client/runtime/model-source.d.ts +19 -2
- package/src/client/runtime/model-source.js +45 -5
- package/src/config/conversion/gemma3/translategemma-4b-1b-enes-q4k-ehf16-af32.json +21 -5
- package/src/config/conversion/gemma3/translategemma-4b-it-q4k-ehf16-af32.json +6 -1
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af16-int4ple.json +1 -1
- package/src/config/conversion/qwen3/qwen-3-reranker-0-6b-q4k-ehf16-af32.json +9 -1
- package/src/config/kernel-path-loader.js +1 -0
- package/src/config/kernels/kernel-ref-digests.js +26 -17
- package/src/config/kernels/registry.json +448 -27
- package/src/config/merge.js +11 -0
- package/src/config/param-validator.js +25 -0
- package/src/config/runtime/default.json +3 -1
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-attn128-candidate.json +20 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-maxbatch64-candidate.json +52 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-normed-ffn-candidate.json +18 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-overlap-ringtokens1-candidate.json +52 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-parity.json +85 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-sequential-candidate.json +52 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-throughput.json +85 -0
- package/src/config/runtime/profiles/qwen-3-5-2b-f16-primary-throughput.json +23 -0
- package/src/config/runtime/profiles/qwen-3-5-2b-metal-parity.json +38 -0
- package/src/config/runtime/profiles/qwen-3-5-2b-metal-throughput.json +135 -0
- package/src/config/runtime/profiles/qwen-3-embedding-0-6b-metal-stability.json +37 -0
- package/src/config/runtime/profiles/qwen-3-embedding-0-6b-metal-throughput.json +21 -0
- package/src/config/runtime/profiles/qwen-3-reranker-0-6b-metal-stability.json +41 -0
- package/src/config/runtime/profiles/qwen-3-reranker-0-6b-metal-throughput.json +41 -0
- package/src/config/runtime/profiles/rerank-stability.json +43 -0
- package/src/config/runtime/profiles/rerank-throughput.json +40 -0
- package/src/config/runtime/profiles/translategemma-1b-nativekd2-q4k-rdna3-throughput-probe.json +22 -0
- package/src/config/schema/doppler.schema.js +40 -5
- package/src/config/schema/execution-v1.schema.d.ts +41 -0
- package/src/config/schema/execution-v1.schema.js +10 -0
- package/src/config/schema/index.d.ts +3 -1
- package/src/config/schema/speculation-self.schema.d.ts +16 -0
- package/src/config/transforms/capability-transform-resolver.js +7 -1
- package/src/config/transforms/execution-graph-transforms.js +2 -2
- package/src/debug/config.d.ts +4 -0
- package/src/experimental/orchestration/index.d.ts +10 -1
- package/src/experimental/training/suite.js +1 -1
- package/src/formats/rdrr/classification.js +3 -0
- package/src/formats/rdrr/integrity.d.ts +0 -2
- package/src/formats/rdrr/integrity.js +0 -2
- package/src/formats/rdrr/types.d.ts +1 -1
- package/src/gpu/kernels/attention.d.ts +2 -0
- package/src/gpu/kernels/attention.js +75 -16
- package/src/gpu/kernels/attention_decode_online_head256_f16kv_output_gate.wgsl +268 -0
- package/src/gpu/kernels/attention_head128_f16kv.wgsl +269 -0
- package/src/gpu/kernels/fused_ffn.d.ts +32 -0
- package/src/gpu/kernels/fused_ffn.js +540 -19
- package/src/gpu/kernels/fused_ffn.wgsl +2 -1
- package/src/gpu/kernels/fused_ffn_f16.wgsl +2 -1
- package/src/gpu/kernels/fused_ffn_f16w_f32a_batched.wgsl +135 -0
- package/src/gpu/kernels/fused_ffn_q4k.wgsl +5 -3
- package/src/gpu/kernels/fused_ffn_q4k_f16.wgsl +2 -1
- package/src/gpu/kernels/fused_ffn_q4k_metal_simd16.wgsl +183 -0
- package/src/gpu/kernels/fused_gate_up_gelu_gated_f16.wgsl +2 -1
- package/src/gpu/kernels/fused_matmul_q4.wgsl +15 -6
- package/src/gpu/kernels/fused_matmul_q4_metal_simd16.wgsl +144 -0
- package/src/gpu/kernels/index.d.ts +21 -0
- package/src/gpu/kernels/index.js +13 -0
- package/src/gpu/kernels/linear-attention-core.js +413 -15
- package/src/gpu/kernels/lm-head-argmax.d.ts +10 -1
- package/src/gpu/kernels/lm-head-argmax.js +153 -57
- package/src/gpu/kernels/lm-head-select-logits.d.ts +23 -0
- package/src/gpu/kernels/lm-head-select-logits.js +186 -0
- package/src/gpu/kernels/lm_head_argmax_q4k.wgsl +289 -0
- package/src/gpu/kernels/lm_head_select_logits.wgsl +77 -0
- package/src/gpu/kernels/matmul-selection.js +5 -1
- package/src/gpu/kernels/residual_rmsnorm_pair.wgsl +127 -0
- package/src/gpu/kernels/rmsnorm-pair.d.ts +33 -0
- package/src/gpu/kernels/rmsnorm-pair.js +113 -0
- package/src/gpu/kernels/rmsnorm-stats.d.ts +29 -0
- package/src/gpu/kernels/rmsnorm-stats.js +250 -0
- package/src/gpu/kernels/silu.wgsl +7 -1
- package/src/gpu/kernels/silu_f16.wgsl +7 -1
- package/src/gpu/kernels/split-qkv-rmsnorm-rope-qk.js +15 -10
- package/src/gpu/kernels/split_qkv_rmsnorm_rope_qk.wgsl +37 -7
- package/src/gpu/kernels/split_qkv_rmsnorm_rope_qk_f16kv_cache.wgsl +37 -7
- package/src/gpu/tensor.d.ts +1 -0
- package/src/index.d.ts +1 -1
- package/src/index.js +1 -1
- package/src/inference/browser-harness-suite-helpers.js +1 -1
- package/src/inference/browser-harness-text-helpers.js +251 -20
- package/src/inference/browser-harness.js +221 -3
- package/src/inference/kv-cache/base.js +42 -1
- package/src/inference/pipelines/structured/json-grammar-mask.d.ts +18 -0
- package/src/inference/pipelines/text/attention/dispatch-params.js +8 -3
- package/src/inference/pipelines/text/attention/output-gate-fusion.js +37 -0
- package/src/inference/pipelines/text/attention/projections.js +174 -4
- package/src/inference/pipelines/text/attention/record.js +44 -21
- package/src/inference/pipelines/text/attention/run.js +18 -3
- package/src/inference/pipelines/text/config.d.ts +18 -0
- package/src/inference/pipelines/text/config.js +100 -0
- package/src/inference/pipelines/text/execution-plan.d.ts +2 -0
- package/src/inference/pipelines/text/execution-plan.js +37 -0
- package/src/inference/pipelines/text/execution-v1.d.ts +6 -0
- package/src/inference/pipelines/text/execution-v1.js +122 -8
- package/src/inference/pipelines/text/ffn/dense.js +239 -21
- package/src/inference/pipelines/text/ffn/standard.js +121 -38
- package/src/inference/pipelines/text/generator-decode-policy.js +1 -0
- package/src/inference/pipelines/text/generator-helpers.js +2 -0
- package/src/inference/pipelines/text/generator-runtime.js +1 -0
- package/src/inference/pipelines/text/generator-steps.js +2 -2
- package/src/inference/pipelines/text/generator.d.ts +18 -1
- package/src/inference/pipelines/text/generator.js +297 -19
- package/src/inference/pipelines/text/init-qkv-fusion.js +155 -62
- package/src/inference/pipelines/text/init.d.ts +22 -3
- package/src/inference/pipelines/text/init.js +2 -2
- package/src/inference/pipelines/text/layer.js +87 -8
- package/src/inference/pipelines/text/linear-attention-ab-fusion.js +208 -0
- package/src/inference/pipelines/text/linear-attention.d.ts +1 -0
- package/src/inference/pipelines/text/linear-attention.js +173 -16
- package/src/inference/pipelines/text/logits/gpu.js +2 -2
- package/src/inference/pipelines/text/logits/index.js +62 -1
- package/src/inference/pipelines/text/model-load.js +6 -0
- package/src/inference/pipelines/text/ops.js +22 -0
- package/src/inference/pipelines/text/per-layer-inputs.js +80 -11
- package/src/inference/pipelines/text/probes.js +9 -4
- package/src/inference/pipelines/text/state.d.ts +4 -4
- package/src/inference/pipelines/text/state.js +4 -0
- package/src/inference/pipelines/text/types.d.ts +27 -2
- package/src/inference/pipelines/text.d.ts +21 -1
- package/src/inference/pipelines/text.js +40 -14
- package/src/inference/tokenizers/bundled.js +45 -23
- package/src/inference/tokenizers/types.d.ts +7 -4
- package/src/loader/doppler-loader.js +25 -2
- package/src/loader/index.d.ts +3 -1
- package/src/loader/index.js +2 -0
- package/src/loader/loader-types.d.ts +3 -1
- package/src/loader/manifest-config.d.ts +18 -0
- package/src/loader/manifest-config.js +11 -3
- package/src/loader/shard-cache.d.ts +2 -0
- package/src/loader/shard-cache.js +29 -1
- package/src/models/gemma4.js +8 -0
- package/src/rules/inference/capability-transforms.rules.json +145 -20
- package/src/rules/inference/ffn.rules.json +33 -0
- package/src/rules/kernels/fused-ffn.rules.json +1 -1
- package/src/rules/kernels/lm-head-argmax.rules.json +10 -0
- package/src/rules/kernels/matmul.rules.json +1 -0
- package/src/rules/kernels/silu.rules.json +1 -0
- package/src/rules/rule-registry.js +2 -0
- package/src/storage/artifact-storage-context.js +36 -0
- package/src/tooling/node-browser-command-runner.js +42 -7
- package/src/tooling/opfs-cache.d.ts +21 -1
- package/src/tooling/opfs-cache.js +168 -32
- package/src/tooling-exports/storage.d.ts +1 -1
- package/src/tooling-exports/storage.js +1 -1
- package/src/types/gpu.d.ts +25 -0
- package/src/version.js +1 -1
- package/tests/kernels/browser/kernel-suite.js +1593 -0
- package/tests/kernels/browser/test-page.js +2629 -0
- package/tests/kernels/harness/benchmark.js +225 -0
- package/tests/kernels/harness/buffer-utils.js +88 -0
- package/tests/kernels/harness/tolerance.js +174 -0
- package/tests/kernels/reference/adam.js +15 -0
- package/tests/kernels/reference/attention.js +158 -0
- package/tests/kernels/reference/backward.js +96 -0
- package/tests/kernels/reference/bias-add.js +20 -0
- package/tests/kernels/reference/conv2d.js +58 -0
- package/tests/kernels/reference/cross-entropy.js +26 -0
- package/tests/kernels/reference/dequant.js +312 -0
- package/tests/kernels/reference/embed.js +15 -0
- package/tests/kernels/reference/gather.js +61 -0
- package/tests/kernels/reference/gelu.js +37 -0
- package/tests/kernels/reference/groupnorm.js +36 -0
- package/tests/kernels/reference/index.js +89 -0
- package/tests/kernels/reference/layernorm.js +22 -0
- package/tests/kernels/reference/matmul.js +57 -0
- package/tests/kernels/reference/modulate.js +24 -0
- package/tests/kernels/reference/moe-gather.js +83 -0
- package/tests/kernels/reference/pixel-shuffle.js +20 -0
- package/tests/kernels/reference/residual.js +55 -0
- package/tests/kernels/reference/rmsnorm.js +52 -0
- package/tests/kernels/reference/rope.js +79 -0
- package/tests/kernels/reference/sample.js +97 -0
- package/tests/kernels/reference/scatter-add.js +58 -0
- package/tests/kernels/reference/silu.js +52 -0
- package/tests/kernels/reference/softmax.js +86 -0
- package/tests/kernels/reference/split-qkv.js +58 -0
- package/tests/kernels/reference/topk.js +88 -0
- package/tests/kernels/reference/upsample2d.js +37 -0
- package/BRANDING.md +0 -14
- package/SECURITY.md +0 -19
- package/assets/doppler-runtime-map.svg +0 -82
- package/src/bootstrap.js +0 -30
- package/src/cli/doppler-cli.d.ts +0 -46
- package/src/cli/doppler-quickstart.d.ts +0 -45
- package/src/cli/doppler-serve.d.ts +0 -210
- package/src/cli/runtime-profiles.d.ts +0 -40
- package/src/client/doppler-api.browser.d.ts +0 -1
- package/src/client/doppler-api.browser.js +0 -25
- package/src/client/doppler-registry.d.ts +0 -54
- package/src/client/runtime/node-quickstart-cache.d.ts +0 -22
- package/src/client/runtime/source-runtime.d.ts +0 -26
- package/src/client/runtime/storage.d.ts +0 -5
- package/src/config/README.md +0 -61
- package/src/config/conversion/README.md +0 -62
- package/src/config/goal-completion-matrix.json +0 -552
- package/src/config/kernels/kernel-ref-digests.d.ts +0 -1
- package/src/config/kernels/moe/gpt-oss.paths.json +0 -49
- package/src/config/kernels/moe/mixtral.paths.json +0 -46
- package/src/config/kernels/registry.d.ts +0 -51
- package/src/config/merge-contract-check.d.ts +0 -16
- package/src/config/merge-contract-check.js +0 -319
- package/src/config/merge-helpers.d.ts +0 -60
- package/src/config/model-lanes/catalog-lane-resolver.d.ts +0 -66
- package/src/config/model-lanes/catalog-lane-resolver.js +0 -170
- package/src/config/param-categories.d.ts +0 -8
- package/src/config/param-validator.d.ts +0 -33
- package/src/config/platforms/loader.d.ts +0 -38
- package/src/config/quantization-contract-check.d.ts +0 -12
- package/src/config/quantization-contract-check.js +0 -91
- package/src/config/runtime/README.md +0 -28
- package/src/config/runtime-merge.d.ts +0 -5
- package/src/config/schema/claim-evidence-contract.schema.json +0 -93
- package/src/config/schema/command-surface-contract.schema.json +0 -81
- package/src/config/schema/goal-completion-matrix.schema.json +0 -223
- package/src/config/schema/model-artifact-contract.schema.json +0 -119
- package/src/config/schema/platform.schema.js +0 -1
- package/src/config/schema/policy-schema-registry.json +0 -31
- package/src/config/schema/subsystem-support-contract.schema.json +0 -48
- package/src/config/support-tiers/subsystems.json +0 -502
- package/src/config/supported-operations.d.ts +0 -14
- package/src/config/transforms/capability-transform-resolver.d.ts +0 -56
- package/src/config/transforms/execution-graph-transforms.d.ts +0 -306
- package/src/converter/conversion-plan.d.ts +0 -56
- package/src/converter/index.d.ts +0 -98
- package/src/converter/index.js +0 -59
- package/src/converter/parsers/diffusion.d.ts +0 -50
- package/src/converter/parsers/gguf.d.ts +0 -31
- package/src/converter/parsers/index.d.ts +0 -21
- package/src/converter/parsers/index.js +0 -12
- package/src/converter/parsers/transformer.d.ts +0 -31
- package/src/converter/quantization-info.d.ts +0 -38
- package/src/converter/rope-config.d.ts +0 -16
- package/src/converter/tokenizer-utils.d.ts +0 -12
- package/src/debug/reference/README.md +0 -114
- package/src/debug/reference/hf_attn_debug.py +0 -114
- package/src/debug/reference/hf_embed_check.py +0 -89
- package/src/debug/reference/hf_layer_out.py +0 -100
- package/src/debug/reference/hf_qwen35_linear_attn_debug.py +0 -268
- package/src/debug/reference/hf_rope_check.py +0 -116
- package/src/debug/reference/hf_weights.py +0 -75
- package/src/errors/doppler-error.d.ts +0 -20
- package/src/experimental/bridge/extension/background.d.ts +0 -14
- package/src/experimental/bridge/extension/background.js +0 -168
- package/src/experimental/bridge/extension/manifest.json +0 -34
- package/src/experimental/bridge/extension-client.d.ts +0 -114
- package/src/experimental/bridge/index.d.ts +0 -69
- package/src/experimental/bridge/node-bridge.js +0 -24
- package/src/experimental/bridge/protocol.d.ts +0 -96
- package/src/experimental/browser/gguf-importer.d.ts +0 -136
- package/src/experimental/browser/gguf-importer.js +0 -532
- package/src/experimental/browser/gguf-parser-browser.d.ts +0 -14
- package/src/experimental/browser/quantization.d.ts +0 -74
- package/src/experimental/browser/safetensors-parser-browser.d.ts +0 -193
- package/src/experimental/browser/shard-io-browser.d.ts +0 -57
- package/src/experimental/browser/shard-io-browser.js +0 -89
- package/src/experimental/browser/tensor-source-download.d.ts +0 -27
- package/src/experimental/diffusion/config/expansion-lanes.json +0 -29
- package/src/experimental/distribution/p2p-transport-contract.d.ts +0 -57
- package/src/experimental/distribution/p2p-transport.js +0 -75
- package/src/experimental/hotswap/intent-bundle.d.ts +0 -37
- package/src/experimental/hotswap/manifest.d.ts +0 -42
- package/src/experimental/hotswap/model-swap.js +0 -42
- package/src/experimental/hotswap/runtime.d.ts +0 -31
- package/src/experimental/training/README.md +0 -164
- package/src/experimental/training/artifacts.d.ts +0 -160
- package/src/experimental/training/attention-backward.d.ts +0 -30
- package/src/experimental/training/attention-forward.d.ts +0 -22
- package/src/experimental/training/attention-forward.js +0 -82
- package/src/experimental/training/autograd.d.ts +0 -51
- package/src/experimental/training/backward-kernels.js +0 -39
- package/src/experimental/training/checkpoint-watch.d.ts +0 -8
- package/src/experimental/training/checkpoint.d.ts +0 -36
- package/src/experimental/training/clip.d.ts +0 -9
- package/src/experimental/training/dataloader.d.ts +0 -8
- package/src/experimental/training/datasets/gepa-frontier.d.ts +0 -41
- package/src/experimental/training/datasets/gepa-frontier.js +0 -220
- package/src/experimental/training/datasets/index.d.ts +0 -24
- package/src/experimental/training/datasets/index.js +0 -13
- package/src/experimental/training/datasets/jsonl.d.ts +0 -11
- package/src/experimental/training/datasets/reploid.d.ts +0 -3
- package/src/experimental/training/datasets/reploid.js +0 -36
- package/src/experimental/training/datasets/teacher-traces.d.ts +0 -51
- package/src/experimental/training/datasets/teacher-traces.js +0 -279
- package/src/experimental/training/datasets/text-pairs.d.ts +0 -66
- package/src/experimental/training/datasets/token-batch.d.ts +0 -21
- package/src/experimental/training/datasets/token-batch.js +0 -52
- package/src/experimental/training/datasets/translation-pairs.d.ts +0 -34
- package/src/experimental/training/datasets/translation-pairs.js +0 -49
- package/src/experimental/training/distillation/artifacts.d.ts +0 -71
- package/src/experimental/training/distillation/checkpoint-watch.d.ts +0 -10
- package/src/experimental/training/distillation/dataset.d.ts +0 -59
- package/src/experimental/training/distillation/eval.d.ts +0 -34
- package/src/experimental/training/distillation/index.d.ts +0 -29
- package/src/experimental/training/distillation/runtime.d.ts +0 -20
- package/src/experimental/training/distillation/scoreboard.d.ts +0 -6
- package/src/experimental/training/distillation/stage-a.d.ts +0 -45
- package/src/experimental/training/distillation/stage-b.d.ts +0 -24
- package/src/experimental/training/distillation/student-fixture.d.ts +0 -27
- package/src/experimental/training/distillation/suite-data.d.ts +0 -45
- package/src/experimental/training/export.d.ts +0 -41
- package/src/experimental/training/index.d.ts +0 -74
- package/src/experimental/training/index.js +0 -63
- package/src/experimental/training/lora-pipeline.d.ts +0 -183
- package/src/experimental/training/lora.d.ts +0 -19
- package/src/experimental/training/loss-scaling.d.ts +0 -21
- package/src/experimental/training/loss.d.ts +0 -10
- package/src/experimental/training/objectives/base.d.ts +0 -58
- package/src/experimental/training/objectives/cross_entropy.d.ts +0 -18
- package/src/experimental/training/objectives/distill_kd.d.ts +0 -16
- package/src/experimental/training/objectives/distill_triplet.d.ts +0 -16
- package/src/experimental/training/objectives/index.d.ts +0 -12
- package/src/experimental/training/objectives/index.js +0 -6
- package/src/experimental/training/objectives/ul_stage1_joint.d.ts +0 -16
- package/src/experimental/training/objectives/ul_stage2_base.d.ts +0 -16
- package/src/experimental/training/operator-agent-eval.d.ts +0 -65
- package/src/experimental/training/operator-artifacts.d.ts +0 -62
- package/src/experimental/training/operator-command.d.ts +0 -5
- package/src/experimental/training/operator-eval.d.ts +0 -48
- package/src/experimental/training/operator-scoreboard.d.ts +0 -5
- package/src/experimental/training/optimizer.d.ts +0 -22
- package/src/experimental/training/runner.d.ts +0 -249
- package/src/experimental/training/suite.d.ts +0 -299
- package/src/experimental/training/tensor-factory.d.ts +0 -9
- package/src/experimental/training/trainer.d.ts +0 -89
- package/src/experimental/training/ul_dataset.d.ts +0 -47
- package/src/experimental/training/ul_schedule.d.ts +0 -6
- package/src/experimental/training/workload-packs/distill-glm52-doppler-agent-harness-gemma270m-f16-sft-tiny.json +0 -160
- package/src/experimental/training/workload-packs/distill-glm52-doppler-code-agent-gemma270m-f16-sft-tiny.json +0 -175
- package/src/experimental/training/workload-packs/distill-glm52-doppler-code-agent-sft-tiny.json +0 -175
- package/src/experimental/training/workload-packs/distill-glm52-doppler-js-json-gemma270m-f16-sft-tiny.json +0 -160
- package/src/experimental/training/workload-packs/distill-glm52-doppler-review-gemma270m-f16-sft-tiny.json +0 -162
- package/src/experimental/training/workload-packs/distill-glm52-doppler-wgsl-gemma270m-f16-sft-tiny.json +0 -160
- package/src/experimental/training/workload-packs/distill-translategemma-medium.json +0 -103
- package/src/experimental/training/workload-packs/distill-translategemma-tiny.json +0 -103
- package/src/experimental/training/workload-packs/lora-doppler-agent-harness-gemma270m-f16-tiny.json +0 -137
- package/src/experimental/training/workload-packs/lora-doppler-code-agent-gemma270m-f16-tiny.json +0 -152
- package/src/experimental/training/workload-packs/lora-doppler-code-agent-tiny.json +0 -152
- package/src/experimental/training/workload-packs/lora-doppler-js-json-gemma270m-f16-tiny.json +0 -137
- package/src/experimental/training/workload-packs/lora-doppler-review-gemma270m-f16-tiny.json +0 -139
- package/src/experimental/training/workload-packs/lora-doppler-wgsl-gemma270m-f16-tiny.json +0 -137
- package/src/experimental/training/workload-packs/lora-toy-medium.json +0 -91
- package/src/experimental/training/workload-packs/lora-toy-tiny.json +0 -91
- package/src/experimental/training/workload-packs/registry.json +0 -278
- package/src/experimental/training/workload-packs/ul-training-medium.json +0 -55
- package/src/experimental/training/workload-packs/ul-training-tiny.json +0 -55
- package/src/experimental/training/workloads.d.ts +0 -201
- package/src/formats/gguf/types.d.ts +0 -74
- package/src/formats/index.d.ts +0 -75
- package/src/formats/index.js +0 -19
- package/src/formats/litert/types.d.ts +0 -74
- package/src/formats/rdrr/distributed/parsing.d.ts +0 -6
- package/src/formats/rdrr/distributed/validation.d.ts +0 -23
- package/src/formats/rdrr/storage-descriptor.d.ts +0 -36
- package/src/formats/rdrr/tensor-config-validator.d.ts +0 -33
- package/src/formats/safetensors/types.d.ts +0 -67
- package/src/formats/tflite/types.d.ts +0 -61
- package/src/formats/tokenizer/index.d.ts +0 -5
- package/src/formats/tokenizer/index.js +0 -3
- package/src/formats/tokenizer/types.d.ts +0 -9
- package/src/gpu/kernel-runtime.d.ts +0 -13
- package/src/gpu/kernel-selection-cache.d.ts +0 -13
- package/src/gpu/kernel-selection-log.d.ts +0 -4
- package/src/gpu/kernel-tuner/benchmarks.d.ts +0 -144
- package/src/gpu/kernel-tuner/cache.d.ts +0 -55
- package/src/gpu/kernel-tuner/index.d.ts +0 -59
- package/src/gpu/kernel-tuner/tuner.d.ts +0 -82
- package/src/gpu/kernel-tuner/types.d.ts +0 -101
- package/src/gpu/kernel-tuner/types.js +0 -4
- package/src/gpu/kernel-tuner.d.ts +0 -33
- package/src/gpu/kernels/README.md +0 -166
- package/src/gpu/kernels/check-finiteness.d.ts +0 -15
- package/src/gpu/kernels/check-hot-vocab-stop.d.ts +0 -18
- package/src/gpu/kernels/check-stop.d.ts +0 -31
- package/src/gpu/kernels/codegen/patches/attention_decode_f16.from.attention_decode.diff +0 -213
- package/src/gpu/kernels/codegen/patches/attention_decode_f16kv.from.attention_decode.diff +0 -89
- package/src/gpu/kernels/codegen/patches/attention_f16.from.attention.diff +0 -62
- package/src/gpu/kernels/codegen/patches/attention_f16kv.from.attention.diff +0 -40
- package/src/gpu/kernels/codegen/patches/attention_small_f16.from.attention_small.diff +0 -226
- package/src/gpu/kernels/codegen/patches/attention_small_f16kv.from.attention_small.diff +0 -264
- package/src/gpu/kernels/codegen/patches/attention_streaming_f16.from.attention_streaming.diff +0 -71
- package/src/gpu/kernels/codegen/patches/attention_streaming_f16kv.from.attention_streaming.diff +0 -53
- package/src/gpu/kernels/codegen/patches/bias_add_f16.from.bias_add.diff +0 -54
- package/src/gpu/kernels/codegen/patches/dequant_mxfp4_expert_f16.from.dequant_mxfp4_expert.diff +0 -125
- package/src/gpu/kernels/codegen/patches/fused_ffn_f16.from.fused_ffn.diff +0 -428
- package/src/gpu/kernels/codegen/patches/fused_matmul_q4_batched_f16.from.fused_matmul_q4_batched.diff +0 -114
- package/src/gpu/kernels/codegen/patches/fused_matmul_rmsnorm_f16.from.fused_matmul_rmsnorm.diff +0 -477
- package/src/gpu/kernels/codegen/patches/fused_swiglu_f16.from.fused_swiglu.diff +0 -60
- package/src/gpu/kernels/codegen/patches/gelu_f16.from.gelu.diff +0 -41
- package/src/gpu/kernels/codegen/patches/layernorm_f16.from.layernorm.diff +0 -82
- package/src/gpu/kernels/codegen/patches/matmul_gemv_residual_f16.from.matmul_gemv_residual.diff +0 -87
- package/src/gpu/kernels/codegen/patches/moe_gather_f16.from.moe_gather.diff +0 -181
- package/src/gpu/kernels/codegen/patches/residual_f16.from.residual.diff +0 -65
- package/src/gpu/kernels/codegen/patches/rmsnorm_f16.from.rmsnorm.diff +0 -515
- package/src/gpu/kernels/codegen/patches/rope_f16.from.rope.diff +0 -292
- package/src/gpu/kernels/codegen/patches/sample_f16.from.sample.diff +0 -336
- package/src/gpu/kernels/codegen/patches/scatter_add_dynamic_f16.from.scatter_add_dynamic.diff +0 -65
- package/src/gpu/kernels/codegen/patches/silu_f16.from.silu.diff +0 -81
- package/src/gpu/kernels/codegen/patches/topk_f16.from.topk.diff +0 -231
- package/src/gpu/kernels/codegen/wgsl-patch-variants.js +0 -152
- package/src/gpu/kernels/codegen/wgsl-variants.js +0 -885
- package/src/gpu/kernels/constants.d.ts +0 -129
- package/src/gpu/kernels/dispatch.d.ts +0 -72
- package/src/gpu/kernels/gated-short-conv.d.ts +0 -63
- package/src/gpu/kernels/kernel-base.d.ts +0 -33
- package/src/gpu/kernels/linear-attention-core.d.ts +0 -42
- package/src/gpu/kernels/matmul-dispatch.d.ts +0 -49
- package/src/gpu/kernels/matmul-selection.d.ts +0 -87
- package/src/gpu/kernels/rep-penalty.d.ts +0 -21
- package/src/gpu/kernels/rule-matcher.d.ts +0 -30
- package/src/gpu/kernels/rule-registry.d.ts +0 -7
- package/src/gpu/kernels/types.js +0 -4
- package/src/gpu/perf-guards.d.ts +0 -13
- package/src/gpu/profiler.d.ts +0 -93
- package/src/gpu/readback-utils.d.ts +0 -16
- package/src/gpu/readback-utils.js +0 -49
- package/src/gpu/submit-probe.d.ts +0 -12
- package/src/index-browser.d.ts +0 -2
- package/src/index-browser.js +0 -2
- package/src/inference/README.md +0 -593
- package/src/inference/browser-harness-contract-helpers.d.ts +0 -5
- package/src/inference/browser-harness-diffusion-energy-suites.d.ts +0 -2
- package/src/inference/browser-harness-model-helpers.d.ts +0 -10
- package/src/inference/browser-harness-report-helpers.d.ts +0 -7
- package/src/inference/browser-harness-suite-helpers.d.ts +0 -40
- package/src/inference/browser-harness-text-helpers.d.ts +0 -152
- package/src/inference/model-lanes/execution-lane-resolver.d.ts +0 -61
- package/src/inference/model-lanes/execution-lane-resolver.js +0 -126
- package/src/inference/network-evolution.js +0 -79
- package/src/inference/pipelines/context.d.ts +0 -21
- package/src/inference/pipelines/diffusion/helpers.d.ts +0 -29
- package/src/inference/pipelines/diffusion/index.d.ts +0 -3
- package/src/inference/pipelines/diffusion/index.js +0 -3
- package/src/inference/pipelines/diffusion/scheduler.d.ts +0 -35
- package/src/inference/pipelines/diffusion/sd3-transformer.d.ts +0 -20
- package/src/inference/pipelines/diffusion/sd3-weights.d.ts +0 -21
- package/src/inference/pipelines/diffusion/text-encoder-gpu.d.ts +0 -82
- package/src/inference/pipelines/diffusion/text-encoder.d.ts +0 -29
- package/src/inference/pipelines/diffusion/types.js +0 -1
- package/src/inference/pipelines/diffusion/vae.d.ts +0 -20
- package/src/inference/pipelines/dream/energy-head-pipeline.d.ts +0 -29
- package/src/inference/pipelines/dream/energy-head-pipeline.js +0 -6
- package/src/inference/pipelines/dream/pipeline.d.ts +0 -17
- package/src/inference/pipelines/dream/pipeline.js +0 -8
- package/src/inference/pipelines/energy/index.d.ts +0 -1
- package/src/inference/pipelines/energy/index.js +0 -1
- package/src/inference/pipelines/energy/types.js +0 -1
- package/src/inference/pipelines/energy-head/index.d.ts +0 -6
- package/src/inference/pipelines/energy-head/index.js +0 -6
- package/src/inference/pipelines/factory.d.ts +0 -10
- package/src/inference/pipelines/index.d.ts +0 -22
- package/src/inference/pipelines/index.js +0 -19
- package/src/inference/pipelines/registry.d.ts +0 -15
- package/src/inference/pipelines/rng.d.ts +0 -2
- package/src/inference/pipelines/structured/tsv-grammar-mask.js +0 -166
- package/src/inference/pipelines/text/attention/attn-config.d.ts +0 -23
- package/src/inference/pipelines/text/attention/dispatch-params.d.ts +0 -173
- package/src/inference/pipelines/text/attention/index.d.ts +0 -24
- package/src/inference/pipelines/text/attention/output-projection.d.ts +0 -12
- package/src/inference/pipelines/text/attention/precision-contract.d.ts +0 -36
- package/src/inference/pipelines/text/attention/projections.d.ts +0 -208
- package/src/inference/pipelines/text/attention/quant-options.d.ts +0 -38
- package/src/inference/pipelines/text/attention/record.d.ts +0 -36
- package/src/inference/pipelines/text/attention/rmsnorm-fusion-gate.d.ts +0 -8
- package/src/inference/pipelines/text/attention/run.d.ts +0 -38
- package/src/inference/pipelines/text/attention/types.d.ts +0 -125
- package/src/inference/pipelines/text/attention.d.ts +0 -23
- package/src/inference/pipelines/text/attention.js +0 -12
- package/src/inference/pipelines/text/bdpa-steamroller.d.ts +0 -22
- package/src/inference/pipelines/text/buffer-types.js +0 -4
- package/src/inference/pipelines/text/debug-utils/config.d.ts +0 -144
- package/src/inference/pipelines/text/debug-utils/index.d.ts +0 -53
- package/src/inference/pipelines/text/debug-utils/logging.d.ts +0 -106
- package/src/inference/pipelines/text/debug-utils/tensor.d.ts +0 -119
- package/src/inference/pipelines/text/debug-utils/utils.d.ts +0 -77
- package/src/inference/pipelines/text/debug-utils.d.ts +0 -42
- package/src/inference/pipelines/text/drift-policy.d.ts +0 -57
- package/src/inference/pipelines/text/dtype-contract.d.ts +0 -53
- package/src/inference/pipelines/text/embed.d.ts +0 -69
- package/src/inference/pipelines/text/execution-runtime-builders.d.ts +0 -49
- package/src/inference/pipelines/text/failure-reduction.d.ts +0 -84
- package/src/inference/pipelines/text/failure-reduction.js +0 -180
- package/src/inference/pipelines/text/ffn/dense.d.ts +0 -99
- package/src/inference/pipelines/text/ffn/index.d.ts +0 -23
- package/src/inference/pipelines/text/ffn/moe.d.ts +0 -24
- package/src/inference/pipelines/text/ffn/sandwich.d.ts +0 -27
- package/src/inference/pipelines/text/ffn/standard.d.ts +0 -26
- package/src/inference/pipelines/text/ffn/types.d.ts +0 -30
- package/src/inference/pipelines/text/ffn.d.ts +0 -31
- package/src/inference/pipelines/text/ffn.js +0 -18
- package/src/inference/pipelines/text/finiteness-guard-status.d.ts +0 -11
- package/src/inference/pipelines/text/finiteness-policy.d.ts +0 -39
- package/src/inference/pipelines/text/generator-decode-policy.d.ts +0 -81
- package/src/inference/pipelines/text/generator-helpers.d.ts +0 -45
- package/src/inference/pipelines/text/generator-prefill-helpers.d.ts +0 -84
- package/src/inference/pipelines/text/generator-runtime.d.ts +0 -93
- package/src/inference/pipelines/text/generator-steps.d.ts +0 -186
- package/src/inference/pipelines/text/index.d.ts +0 -5
- package/src/inference/pipelines/text/index.js +0 -6
- package/src/inference/pipelines/text/kernel-trace.d.ts +0 -119
- package/src/inference/pipelines/text/layer.d.ts +0 -118
- package/src/inference/pipelines/text/load-errors.d.ts +0 -21
- package/src/inference/pipelines/text/logits/precision-policy.d.ts +0 -11
- package/src/inference/pipelines/text/logits/types.js +0 -4
- package/src/inference/pipelines/text/logits.d.ts +0 -27
- package/src/inference/pipelines/text/logits.js +0 -16
- package/src/inference/pipelines/text/lora-apply.d.ts +0 -28
- package/src/inference/pipelines/text/moe-cache.d.ts +0 -32
- package/src/inference/pipelines/text/moe-cpu-gptoss.d.ts +0 -9
- package/src/inference/pipelines/text/moe-cpu.d.ts +0 -13
- package/src/inference/pipelines/text/moe-gpu.d.ts +0 -26
- package/src/inference/pipelines/text/moe-helpers.d.ts +0 -12
- package/src/inference/pipelines/text/moe-shape-validator.d.ts +0 -70
- package/src/inference/pipelines/text/operator-events.d.ts +0 -125
- package/src/inference/pipelines/text/operator-identity.d.ts +0 -60
- package/src/inference/pipelines/text/ops.d.ts +0 -222
- package/src/inference/pipelines/text/per-layer-inputs.d.ts +0 -107
- package/src/inference/pipelines/text/probes.d.ts +0 -28
- package/src/inference/pipelines/text/sampling-config.d.ts +0 -48
- package/src/inference/pipelines/text/stage-names.d.ts +0 -88
- package/src/inference/pipelines/text/tsir-fixture-writer.d.ts +0 -57
- package/src/inference/pipelines/text/types.js +0 -4
- package/src/inference/tokenizers/base.d.ts +0 -44
- package/src/inference/tokenizers/behavior-flags.d.ts +0 -9
- package/src/inference/tokenizers/bpe.d.ts +0 -27
- package/src/inference/tokenizers/bundled.d.ts +0 -67
- package/src/inference/tokenizers/sentencepiece.d.ts +0 -30
- package/src/inference/tokenizers/types.js +0 -7
- package/src/loader/embedding-limit-preflight.d.ts +0 -31
- package/src/loader/per-layer-input-loader.d.ts +0 -30
- package/src/loader/tensors/functional-descriptor-loader.d.ts +0 -9
- package/src/loader/tensors/source-transform.d.ts +0 -65
- package/src/memory/address-table.d.ts +0 -96
- package/src/rules/execution-rules-contract-check.d.ts +0 -21
- package/src/rules/layer-pattern-contract-check.d.ts +0 -17
- package/src/rules/rule-registry.d.ts +0 -77
- package/src/storage/artifact-storage-context.d.ts +0 -63
- package/src/storage/backends/idb-store.d.ts +0 -52
- package/src/storage/backends/memory-store.d.ts +0 -36
- package/src/storage/backends/opfs-store.d.ts +0 -41
- package/src/storage/blake3.d.ts +0 -17
- package/src/storage/download-types.d.ts +0 -157
- package/src/storage/downloader.d.ts +0 -103
- package/src/storage/index.d.ts +0 -274
- package/src/storage/index.js +0 -201
- package/src/storage/preflight.d.ts +0 -144
- package/src/storage/preflight.js +0 -360
- package/src/storage/quickstart-downloader.d.ts +0 -157
- package/src/storage/quickstart-downloader.js +0 -289
- package/src/storage/source-artifact-store.d.ts +0 -59
- package/src/sw.js +0 -184
- package/src/tooling/command-api-constants.d.ts +0 -9
- package/src/tooling/command-api-family-normalizers.d.ts +0 -10
- package/src/tooling/command-api-helpers.d.ts +0 -23
- package/src/tooling/command-runner-shared.d.ts +0 -89
- package/src/tooling/conversion-config-materializer.d.ts +0 -14
- package/src/tooling/conversion-config-materializer.js +0 -96
- package/src/tooling/diagnose-runner.d.ts +0 -28
- package/src/tooling/hf-registry-utils.d.ts +0 -55
- package/src/tooling/kernel-path-builder/custom-runtime-facts.d.ts +0 -18
- package/src/tooling/kernel-path-builder/custom-runtime-facts.js +0 -72
- package/src/tooling/kernel-path-builder/index.d.ts +0 -65
- package/src/tooling/kernel-path-builder/index.js +0 -1361
- package/src/tooling/kernel-path-builder/runtime-overlay.d.ts +0 -9
- package/src/tooling/kernel-path-builder/runtime-overlay.js +0 -147
- package/src/tooling/lean-execution-contract-runner.d.ts +0 -33
- package/src/tooling/lean-execution-contract-runner.js +0 -158
- package/src/tooling/lean-execution-contract.d.ts +0 -16
- package/src/tooling/lean-execution-contract.js +0 -81
- package/src/tooling/litert-package-runtime.d.ts +0 -46
- package/src/tooling/node-convert-worker-pool.d.ts +0 -16
- package/src/tooling/node-convert-worker.d.ts +0 -1
- package/src/tooling/node-converter.d.ts +0 -1
- package/src/tooling/node-file-fetch.d.ts +0 -1
- package/src/tooling/node-source-runtime.d.ts +0 -24
- package/src/tooling/node-webgpu.d.ts +0 -17
- package/src/tooling/precision-replay-math.js +0 -174
- package/src/tooling/program-bundle-host.d.ts +0 -16
- package/src/tooling/program-bundle-host.js +0 -20
- package/src/tooling/runtime-input-composition.d.ts +0 -37
- package/src/tooling/source-package-profiles.d.ts +0 -50
- package/src/tooling/source-runtime-converter-config.d.ts +0 -19
- package/src/tooling/source-runtime-materializer.d.ts +0 -6
- package/src/tooling/source-runtime-materializer.js +0 -89
- package/src/tooling-exports.browser.d.ts +0 -7
- package/src/types/chrome.d.ts +0 -36
- package/src/types/chrome.js +0 -1
- package/src/types/index.d.ts +0 -1
- package/src/types/index.js +0 -1
- package/src/utils/array-buffer.d.ts +0 -4
- package/src/utils/canonical-hash.d.ts +0 -7
- package/src/utils/clone-json.d.ts +0 -1
- package/src/utils/hf-resolve-url.d.ts +0 -16
- package/src/utils/index.d.ts +0 -10
- package/src/utils/index.js +0 -10
- package/src/utils/load-json.d.ts +0 -5
- package/src/utils/plain-object.d.ts +0 -1
- package/src/utils/sha256.d.ts +0 -5
- package/src/utils/stable-sort-object.d.ts +0 -1
|
@@ -0,0 +1,37 @@
|
|
|
1
|
+
{
|
|
2
|
+
"id": "profiles/qwen-3-embedding-0-6b-metal-stability",
|
|
3
|
+
"name": "qwen-3-embedding-0-6b-metal-stability",
|
|
4
|
+
"description": "Qwen 3 Embedding 0.6B Metal validation profile preserving manifest compute precision.",
|
|
5
|
+
"intent": "verify",
|
|
6
|
+
"stability": "canonical",
|
|
7
|
+
"owner": "doppler-core",
|
|
8
|
+
"createdAtUtc": "2026-07-09T00:00:00Z",
|
|
9
|
+
"extends": "profiles/default",
|
|
10
|
+
"runtime": {
|
|
11
|
+
"shared": {
|
|
12
|
+
"tooling": {
|
|
13
|
+
"intent": "verify"
|
|
14
|
+
},
|
|
15
|
+
"debug": {
|
|
16
|
+
"logLevel": {
|
|
17
|
+
"defaultLogLevel": "silent"
|
|
18
|
+
},
|
|
19
|
+
"trace": {
|
|
20
|
+
"enabled": false
|
|
21
|
+
}
|
|
22
|
+
}
|
|
23
|
+
},
|
|
24
|
+
"inference": {
|
|
25
|
+
"batching": {
|
|
26
|
+
"maxTokens": 1
|
|
27
|
+
},
|
|
28
|
+
"sampling": {
|
|
29
|
+
"temperature": 0
|
|
30
|
+
},
|
|
31
|
+
"session": {
|
|
32
|
+
"prefillChunkSubmitMode": "async",
|
|
33
|
+
"skipEmbeddingKVCacheWrites": true
|
|
34
|
+
}
|
|
35
|
+
}
|
|
36
|
+
}
|
|
37
|
+
}
|
|
@@ -0,0 +1,21 @@
|
|
|
1
|
+
{
|
|
2
|
+
"id": "profiles/qwen-3-embedding-0-6b-metal-throughput",
|
|
3
|
+
"name": "qwen-3-embedding-0-6b-metal-throughput",
|
|
4
|
+
"description": "Qwen 3 Embedding 0.6B Metal throughput profile preserving manifest compute precision.",
|
|
5
|
+
"intent": "calibrate",
|
|
6
|
+
"stability": "canonical",
|
|
7
|
+
"owner": "doppler-core",
|
|
8
|
+
"createdAtUtc": "2026-07-09T00:00:00Z",
|
|
9
|
+
"extends": "profiles/throughput",
|
|
10
|
+
"runtime": {
|
|
11
|
+
"inference": {
|
|
12
|
+
"batching": {
|
|
13
|
+
"maxTokens": 1
|
|
14
|
+
},
|
|
15
|
+
"session": {
|
|
16
|
+
"prefillChunkSubmitMode": "async",
|
|
17
|
+
"skipEmbeddingKVCacheWrites": true
|
|
18
|
+
}
|
|
19
|
+
}
|
|
20
|
+
}
|
|
21
|
+
}
|
|
@@ -0,0 +1,41 @@
|
|
|
1
|
+
{
|
|
2
|
+
"id": "profiles/qwen-3-reranker-0-6b-metal-stability",
|
|
3
|
+
"name": "qwen-3-reranker-0-6b-metal-stability",
|
|
4
|
+
"description": "Qwen 3 Reranker 0.6B browser/Metal validation profile preserving manifest compute precision while exercising the fixed head_dim=128 f16-KV prefill attention path.",
|
|
5
|
+
"intent": "verify",
|
|
6
|
+
"stability": "canonical",
|
|
7
|
+
"owner": "doppler-core",
|
|
8
|
+
"createdAtUtc": "2026-07-09T00:00:00Z",
|
|
9
|
+
"extends": "profiles/rerank-stability",
|
|
10
|
+
"runtime": {
|
|
11
|
+
"inference": {
|
|
12
|
+
"executionPatch": {
|
|
13
|
+
"addKernels": [
|
|
14
|
+
{
|
|
15
|
+
"key": "attn_head128",
|
|
16
|
+
"kernel": {
|
|
17
|
+
"kernel": "attention_head128_f16kv.wgsl",
|
|
18
|
+
"entry": "main",
|
|
19
|
+
"digest": "sha256:559a3da4f3f5dcc4671f5a9b429a84209340a66a6e17f583bb329b2e23832c9b",
|
|
20
|
+
"precision": {
|
|
21
|
+
"kvDtype": "f16"
|
|
22
|
+
}
|
|
23
|
+
}
|
|
24
|
+
}
|
|
25
|
+
],
|
|
26
|
+
"set": [
|
|
27
|
+
{
|
|
28
|
+
"section": "prefill",
|
|
29
|
+
"op": "attention",
|
|
30
|
+
"kernelKey": "attn_head128"
|
|
31
|
+
}
|
|
32
|
+
]
|
|
33
|
+
},
|
|
34
|
+
"session": {
|
|
35
|
+
"prefillChunkLayers": 28,
|
|
36
|
+
"retainQ4KMaterialization": true,
|
|
37
|
+
"useWideTileQ4KPrefill": true
|
|
38
|
+
}
|
|
39
|
+
}
|
|
40
|
+
}
|
|
41
|
+
}
|
|
@@ -0,0 +1,41 @@
|
|
|
1
|
+
{
|
|
2
|
+
"id": "profiles/qwen-3-reranker-0-6b-metal-throughput",
|
|
3
|
+
"name": "qwen-3-reranker-0-6b-metal-throughput",
|
|
4
|
+
"description": "Qwen 3 Reranker 0.6B browser/Metal throughput profile preserving manifest compute precision while using fixed head_dim=128 f16-KV prefill attention.",
|
|
5
|
+
"intent": "calibrate",
|
|
6
|
+
"stability": "canonical",
|
|
7
|
+
"owner": "doppler-core",
|
|
8
|
+
"createdAtUtc": "2026-07-09T00:00:00Z",
|
|
9
|
+
"extends": "profiles/rerank-throughput",
|
|
10
|
+
"runtime": {
|
|
11
|
+
"inference": {
|
|
12
|
+
"executionPatch": {
|
|
13
|
+
"addKernels": [
|
|
14
|
+
{
|
|
15
|
+
"key": "attn_head128",
|
|
16
|
+
"kernel": {
|
|
17
|
+
"kernel": "attention_head128_f16kv.wgsl",
|
|
18
|
+
"entry": "main",
|
|
19
|
+
"digest": "sha256:559a3da4f3f5dcc4671f5a9b429a84209340a66a6e17f583bb329b2e23832c9b",
|
|
20
|
+
"precision": {
|
|
21
|
+
"kvDtype": "f16"
|
|
22
|
+
}
|
|
23
|
+
}
|
|
24
|
+
}
|
|
25
|
+
],
|
|
26
|
+
"set": [
|
|
27
|
+
{
|
|
28
|
+
"section": "prefill",
|
|
29
|
+
"op": "attention",
|
|
30
|
+
"kernelKey": "attn_head128"
|
|
31
|
+
}
|
|
32
|
+
]
|
|
33
|
+
},
|
|
34
|
+
"session": {
|
|
35
|
+
"prefillChunkLayers": 28,
|
|
36
|
+
"retainQ4KMaterialization": true,
|
|
37
|
+
"useWideTileQ4KPrefill": true
|
|
38
|
+
}
|
|
39
|
+
}
|
|
40
|
+
}
|
|
41
|
+
}
|
|
@@ -0,0 +1,43 @@
|
|
|
1
|
+
{
|
|
2
|
+
"id": "profiles/rerank-stability",
|
|
3
|
+
"name": "rerank-stability",
|
|
4
|
+
"description": "Stable rerank validation profile with deterministic scoring and quiet tracing.",
|
|
5
|
+
"intent": "verify",
|
|
6
|
+
"stability": "canonical",
|
|
7
|
+
"owner": "doppler-core",
|
|
8
|
+
"createdAtUtc": "2026-07-06T00:00:00Z",
|
|
9
|
+
"extends": "profiles/default",
|
|
10
|
+
"runtime": {
|
|
11
|
+
"shared": {
|
|
12
|
+
"tooling": {
|
|
13
|
+
"intent": "verify"
|
|
14
|
+
},
|
|
15
|
+
"debug": {
|
|
16
|
+
"logLevel": {
|
|
17
|
+
"defaultLogLevel": "silent"
|
|
18
|
+
},
|
|
19
|
+
"trace": {
|
|
20
|
+
"enabled": false
|
|
21
|
+
}
|
|
22
|
+
}
|
|
23
|
+
},
|
|
24
|
+
"inference": {
|
|
25
|
+
"batching": {
|
|
26
|
+
"maxTokens": 1024
|
|
27
|
+
},
|
|
28
|
+
"sampling": {
|
|
29
|
+
"temperature": 0
|
|
30
|
+
},
|
|
31
|
+
"kvcache": {
|
|
32
|
+
"maxSeqLen": 1024,
|
|
33
|
+
"kvDtype": "f16"
|
|
34
|
+
},
|
|
35
|
+
"session": {
|
|
36
|
+
"kvcache": {
|
|
37
|
+
"maxSeqLen": 1024,
|
|
38
|
+
"kvDtype": "f16"
|
|
39
|
+
}
|
|
40
|
+
}
|
|
41
|
+
}
|
|
42
|
+
}
|
|
43
|
+
}
|
|
@@ -0,0 +1,40 @@
|
|
|
1
|
+
{
|
|
2
|
+
"id": "profiles/rerank-throughput",
|
|
3
|
+
"name": "rerank-throughput",
|
|
4
|
+
"description": "Rerank benchmark profile for repeated yes/no-logit scoring throughput runs.",
|
|
5
|
+
"intent": "calibrate",
|
|
6
|
+
"stability": "canonical",
|
|
7
|
+
"owner": "doppler-core",
|
|
8
|
+
"createdAtUtc": "2026-07-06T00:00:00Z",
|
|
9
|
+
"extends": "profiles/rerank-stability",
|
|
10
|
+
"runtime": {
|
|
11
|
+
"shared": {
|
|
12
|
+
"tooling": {
|
|
13
|
+
"intent": "calibrate"
|
|
14
|
+
},
|
|
15
|
+
"debug": {
|
|
16
|
+
"logOutput": {
|
|
17
|
+
"stdout": false
|
|
18
|
+
}
|
|
19
|
+
}
|
|
20
|
+
},
|
|
21
|
+
"inference": {
|
|
22
|
+
"batching": {
|
|
23
|
+
"maxTokens": 256
|
|
24
|
+
},
|
|
25
|
+
"sampling": {
|
|
26
|
+
"temperature": 0
|
|
27
|
+
},
|
|
28
|
+
"kvcache": {
|
|
29
|
+
"maxSeqLen": 256,
|
|
30
|
+
"kvDtype": "f16"
|
|
31
|
+
},
|
|
32
|
+
"session": {
|
|
33
|
+
"kvcache": {
|
|
34
|
+
"maxSeqLen": 256,
|
|
35
|
+
"kvDtype": "f16"
|
|
36
|
+
}
|
|
37
|
+
}
|
|
38
|
+
}
|
|
39
|
+
}
|
|
40
|
+
}
|
package/src/config/runtime/profiles/translategemma-1b-nativekd2-q4k-rdna3-throughput-probe.json
ADDED
|
@@ -0,0 +1,22 @@
|
|
|
1
|
+
{
|
|
2
|
+
"id": "profiles/translategemma-1b-nativekd2-q4k-rdna3-throughput-probe",
|
|
3
|
+
"name": "translategemma-1b-nativekd2-q4k-rdna3-throughput-probe",
|
|
4
|
+
"description": "Experimental TranslateGemma 1B NativeKD2 EN/ES calibration profile for AMD RDNA3/Vulkan evidence. Preserves the manifest 4x1 sequential decode cadence while retaining Q4K materialization, selecting wide-tile Q4K prefill, and enabling exact-safe packed-QKV and RMSNorm-pair fusions.",
|
|
5
|
+
"intent": "calibrate",
|
|
6
|
+
"stability": "experimental",
|
|
7
|
+
"owner": "doppler-core",
|
|
8
|
+
"createdAtUtc": "2026-07-11T00:00:00Z",
|
|
9
|
+
"extends": "profiles/throughput",
|
|
10
|
+
"model": "translategemma-4b-1b-enes-q4k-ehf16-af32",
|
|
11
|
+
"runtime": {
|
|
12
|
+
"inference": {
|
|
13
|
+
"session": {
|
|
14
|
+
"retainQ4KMaterialization": true,
|
|
15
|
+
"useWideTileQ4KPrefill": true,
|
|
16
|
+
"useSandwichRMSNormPairFusion": true,
|
|
17
|
+
"usePostFfnNextInputRMSNormPairFusion": true,
|
|
18
|
+
"useFusedQKVSplitQKNormRoPE": true
|
|
19
|
+
}
|
|
20
|
+
}
|
|
21
|
+
}
|
|
22
|
+
}
|
|
@@ -89,6 +89,7 @@ export const DEFAULT_RUNTIME_CONFIG = {
|
|
|
89
89
|
// memory-conservative; profiles with known-safe memory can raise it.
|
|
90
90
|
prefillChunkLayers: 4,
|
|
91
91
|
prefillTokenChunkSize: null,
|
|
92
|
+
skipEmbeddingKVCacheWrites: false,
|
|
92
93
|
// Opt into the flash-attention prefill kernel (head_dim=256, f16 KV,
|
|
93
94
|
// contiguous layout). Two-pass dispatch raises RDNA3 workgroup
|
|
94
95
|
// occupancy by KV-axis splitting. The kernel itself enforces head_dim
|
|
@@ -101,6 +102,10 @@ export const DEFAULT_RUNTIME_CONFIG = {
|
|
|
101
102
|
// activation + no LoRA on gate/up. Default false until parity is
|
|
102
103
|
// validated on target hardware.
|
|
103
104
|
useFusedGateUpGelu: false,
|
|
105
|
+
// Opt into large-batch f16-weight/f32-activation fused gate/up prefill.
|
|
106
|
+
// This keeps the AF32 lane while replacing split gate/up/activation
|
|
107
|
+
// dispatches for explicitly profiled browser/Metal rerank profiles.
|
|
108
|
+
useLargeBatchF16F32FusedGateUp: false,
|
|
104
109
|
// Opt into the register-tiled Q4_K prefill matmul kernel (64x64 outputs
|
|
105
110
|
// per workgroup via 4x4 register tile per thread). Amortizes Q4_K
|
|
106
111
|
// dequantization across 16x the outputs of q4_fused_batched_f16a,
|
|
@@ -139,6 +144,29 @@ export const DEFAULT_RUNTIME_CONFIG = {
|
|
|
139
144
|
// its normal observation point. Default false until correctness + perf
|
|
140
145
|
// validated.
|
|
141
146
|
usePostFfnNextInputRMSNormPairFusion: false,
|
|
147
|
+
// Opt into a decode-only post-attention norm stats prelude consumed by
|
|
148
|
+
// Q4_K fused gate/up. It materializes the pre-norm residual sum and one
|
|
149
|
+
// inverse-RMS scalar instead of a full normalized FFN input tensor.
|
|
150
|
+
// Default false until correctness + perf validated on target hardware.
|
|
151
|
+
usePostAttnNormFusedGateUp: false,
|
|
152
|
+
// Optional Q4_K fused FFN pipeline constants keyed by phase. Profiles can
|
|
153
|
+
// tune the gate/up workgroup shape per platform without changing the
|
|
154
|
+
// global kernel default or manifest math contract.
|
|
155
|
+
fusedFfnQ4K: null,
|
|
156
|
+
lmHeadArgmaxQ4K: null,
|
|
157
|
+
attentionDecodeOnline: null,
|
|
158
|
+
// Opt into a decode-only linear-attention A+B input projection fusion.
|
|
159
|
+
// Requires dense f16 row weights for linear_attn.in_proj_a/b and falls
|
|
160
|
+
// back to separate projections when the weight contract is not met.
|
|
161
|
+
useLinearAttentionABProjectionFusion: false,
|
|
162
|
+
// Opt into a decode-only linear-attention QKV+Z projection fusion.
|
|
163
|
+
// Requires row-wise Q4_K linear_attn.in_proj_qkv/z weights with matching
|
|
164
|
+
// input width and falls back to separate projections otherwise.
|
|
165
|
+
useLinearAttentionQKVZProjectionFusion: false,
|
|
166
|
+
// Opt into a decode-only linear-attention core fusion. Combines the
|
|
167
|
+
// conv1d state update and recurrent scan when qRep=1 and head dims fit
|
|
168
|
+
// one workgroup, eliminating the conv_out intermediate dispatch.
|
|
169
|
+
useLinearAttentionFusedDecodeCore: false,
|
|
142
170
|
// Opt into the single-pass flash attention prefill kernel adapted from
|
|
143
171
|
// ORT's flash_attention.wgsl.template. 64 threads = 64 queries per WG;
|
|
144
172
|
// private Q/O tiles, shared K/V tiles, online softmax, no reduce pass.
|
|
@@ -279,12 +307,19 @@ function normalizeSessionContracts(inference, base, overrides) {
|
|
|
279
307
|
}
|
|
280
308
|
|
|
281
309
|
const overrideSessionCompute = overrideSession.compute;
|
|
282
|
-
if (isPlainObject(overrideSessionCompute)
|
|
310
|
+
if (isPlainObject(overrideSessionCompute)) {
|
|
283
311
|
const compute = isPlainObject(inference.session.compute) ? inference.session.compute : {};
|
|
284
|
-
|
|
285
|
-
|
|
286
|
-
|
|
287
|
-
|
|
312
|
+
const nextCompute = { ...compute };
|
|
313
|
+
if (hasOwn(overrideSessionCompute, 'defaults')) {
|
|
314
|
+
nextCompute.defaults = replaceSubtree(
|
|
315
|
+
overrideSessionCompute.defaults,
|
|
316
|
+
baseSession.compute?.defaults
|
|
317
|
+
);
|
|
318
|
+
}
|
|
319
|
+
if (hasOwn(overrideSessionCompute, 'kernelProfiles')) {
|
|
320
|
+
nextCompute.kernelProfiles = overrideSessionCompute.kernelProfiles;
|
|
321
|
+
}
|
|
322
|
+
inference.session.compute = nextCompute;
|
|
288
323
|
}
|
|
289
324
|
}
|
|
290
325
|
|
|
@@ -105,6 +105,8 @@ export interface ExecutionV1DecodeLoopSchema {
|
|
|
105
105
|
* Set to `null` to disable auto (treat as sequential).
|
|
106
106
|
*/
|
|
107
107
|
submitLatencyThresholdMs: number | null;
|
|
108
|
+
/** Explicit cap for one multi-token decode submission. Absent uses rule policy. */
|
|
109
|
+
maxBatchDecodeTokens?: number | null;
|
|
108
110
|
ringTokens: number | null;
|
|
109
111
|
ringStop: number | null;
|
|
110
112
|
ringStaging: number | null;
|
|
@@ -177,8 +179,12 @@ export interface ExecutionV1SessionSchema {
|
|
|
177
179
|
prefillChunkSubmitMode: 'sync' | 'async';
|
|
178
180
|
/** Nullable token count for streaming prompt prefill before final logits. */
|
|
179
181
|
prefillTokenChunkSize: number | null;
|
|
182
|
+
/** Skip KV-cache writes for embedding-only hidden-state prefill routes. */
|
|
183
|
+
skipEmbeddingKVCacheWrites: boolean;
|
|
180
184
|
/** Opt into flash-attention prefill kernel. Requires head_dim=256, f16 KV, contiguous layout. */
|
|
181
185
|
useFlashPrefillAttention: boolean;
|
|
186
|
+
/** Opt into large-batch f16-weight/f32-activation fused gate/up prefill. */
|
|
187
|
+
useLargeBatchF16F32FusedGateUp: boolean;
|
|
182
188
|
/** Opt into WideTile Q4_K prefill matmul. Requires f32 activations + Q4_K weights + shader-f16 + M>=TILE_M. */
|
|
183
189
|
useWideTileQ4KPrefill: boolean;
|
|
184
190
|
/** Opt into WideTile Q4_K decode matmul. Requires f32 activations + Q4_K weights + shader-f16 + M=1. */
|
|
@@ -187,6 +193,41 @@ export interface ExecutionV1SessionSchema {
|
|
|
187
193
|
useSandwichRMSNormPairFusion: boolean;
|
|
188
194
|
/** Opt into post-FFN plus next-layer input RMSNorm decode fusion. Requires f32 activations + M=1. */
|
|
189
195
|
usePostFfnNextInputRMSNormPairFusion: boolean;
|
|
196
|
+
/** Opt into post-attention RMSNorm stats consumed by Q4_K fused gate/up. */
|
|
197
|
+
usePostAttnNormFusedGateUp: boolean;
|
|
198
|
+
/** Optional Q4_K fused gate/up variant and pipeline constants keyed by phase. */
|
|
199
|
+
fusedFfnQ4K: {
|
|
200
|
+
decode?: {
|
|
201
|
+
variant?: 'q4k_metal_simd16' | null;
|
|
202
|
+
pipelineConstants?: Record<string, number | boolean> | null;
|
|
203
|
+
};
|
|
204
|
+
prefill?: {
|
|
205
|
+
variant?: null;
|
|
206
|
+
pipelineConstants?: Record<string, number | boolean> | null;
|
|
207
|
+
};
|
|
208
|
+
} | null;
|
|
209
|
+
/** Optional Q4_K LM-head argmax tuning policy. */
|
|
210
|
+
lmHeadArgmaxQ4K: {
|
|
211
|
+
useFullBlockFastPath?: boolean;
|
|
212
|
+
colsPerWorkgroup?: number;
|
|
213
|
+
threadsPerCol?: number;
|
|
214
|
+
} | null;
|
|
215
|
+
/** Optional online decode attention pipeline constants. */
|
|
216
|
+
attentionDecodeOnline: {
|
|
217
|
+
workgroupSize?: 128 | 256;
|
|
218
|
+
useDirectContiguousKVLayout?: boolean;
|
|
219
|
+
useOutputGateFusion?: boolean;
|
|
220
|
+
} | null;
|
|
221
|
+
/** Opt into linear-attention A+B projection decode fusion. Requires dense f16 row A/B weights + M=1. */
|
|
222
|
+
useLinearAttentionABProjectionFusion: boolean;
|
|
223
|
+
/** Opt into linear-attention QKV+Z projection decode fusion. Requires row-wise Q4_K QKV/Z weights + M=1. */
|
|
224
|
+
useLinearAttentionQKVZProjectionFusion: boolean;
|
|
225
|
+
/** Opt into linear-attention decode core fusion. Requires qRep=1 and head dims fitting one workgroup. */
|
|
226
|
+
useLinearAttentionFusedDecodeCore: boolean;
|
|
227
|
+
/** Opt into WideTile Q4_K matmul + residual epilogue fusion. */
|
|
228
|
+
useWideTileResidualFusion: boolean;
|
|
229
|
+
/** Opt into RMSNorm + WideTile Q4_K matmul prologue fusion. */
|
|
230
|
+
useFusedRmsnormWideTile: boolean;
|
|
190
231
|
/** Opt into fused packed-QKV split plus weighted Q/K RMSNorm. Requires f32 QKV output and weighted Q/K norm. */
|
|
191
232
|
useFusedQKVSplitQKNorm: boolean;
|
|
192
233
|
/** Opt into fused packed-QKV split plus weighted Q/K RMSNorm and full-head non-interleaved RoPE. */
|
|
@@ -71,11 +71,21 @@ export const DEFAULT_EXECUTION_V1_SESSION = {
|
|
|
71
71
|
speculation: null,
|
|
72
72
|
prefillChunkSubmitMode: 'sync',
|
|
73
73
|
prefillTokenChunkSize: null,
|
|
74
|
+
skipEmbeddingKVCacheWrites: false,
|
|
74
75
|
useFlashPrefillAttention: false,
|
|
76
|
+
useLargeBatchF16F32FusedGateUp: false,
|
|
75
77
|
useWideTileQ4KPrefill: false,
|
|
76
78
|
useWideTileQ4KDecode: false,
|
|
77
79
|
useSandwichRMSNormPairFusion: false,
|
|
78
80
|
usePostFfnNextInputRMSNormPairFusion: false,
|
|
81
|
+
usePostAttnNormFusedGateUp: false,
|
|
82
|
+
lmHeadArgmaxQ4K: null,
|
|
83
|
+
attentionDecodeOnline: null,
|
|
84
|
+
useLinearAttentionABProjectionFusion: false,
|
|
85
|
+
useLinearAttentionQKVZProjectionFusion: false,
|
|
86
|
+
useLinearAttentionFusedDecodeCore: false,
|
|
87
|
+
useWideTileResidualFusion: false,
|
|
88
|
+
useFusedRmsnormWideTile: false,
|
|
79
89
|
useFusedQKVSplitQKNorm: false,
|
|
80
90
|
useFusedQKVSplitQKNormRoPE: false,
|
|
81
91
|
retainQ4KMaterialization: false,
|
|
@@ -96,6 +96,8 @@ export {
|
|
|
96
96
|
type NormalizationSchema,
|
|
97
97
|
type FFNSchema,
|
|
98
98
|
type LayerPipelineOp,
|
|
99
|
+
type LayerPipelinePhase,
|
|
100
|
+
type LayerPipelineDtype,
|
|
99
101
|
type LayerPipelineNormWeight,
|
|
100
102
|
type LayerPipelineStepSchema,
|
|
101
103
|
type LayerPipelineOverrideSchema,
|
|
@@ -902,7 +904,7 @@ export {
|
|
|
902
904
|
// =============================================================================
|
|
903
905
|
// Tuner Limits
|
|
904
906
|
// =============================================================================
|
|
905
|
-
export { DEFAULT_TUNER_LIMITS } from './
|
|
907
|
+
export { DEFAULT_TUNER_LIMITS } from './kernel-thresholds.schema.js';
|
|
906
908
|
|
|
907
909
|
// =============================================================================
|
|
908
910
|
// Self-Speculation Schema
|
|
@@ -0,0 +1,16 @@
|
|
|
1
|
+
export type SelfSpeculationMode = 'none' | 'self' | 'draft' | 'medusa';
|
|
2
|
+
export type SelfSpeculationVerifyMode = 'greedy';
|
|
3
|
+
|
|
4
|
+
export interface SelfSpeculationConfig {
|
|
5
|
+
mode: SelfSpeculationMode;
|
|
6
|
+
tokens: number;
|
|
7
|
+
verify: SelfSpeculationVerifyMode;
|
|
8
|
+
threshold: number | null;
|
|
9
|
+
rollbackOnReject: boolean;
|
|
10
|
+
}
|
|
11
|
+
|
|
12
|
+
export const SPECULATION_MODES: readonly SelfSpeculationMode[];
|
|
13
|
+
export const SPECULATION_VERIFY_MODES: readonly SelfSpeculationVerifyMode[];
|
|
14
|
+
export const DEFAULT_SELF_SPECULATION_CONFIG: Readonly<SelfSpeculationConfig>;
|
|
15
|
+
|
|
16
|
+
export function validateSelfSpeculationConfig(config: SelfSpeculationConfig): void;
|
|
@@ -20,13 +20,16 @@ const rules = await loadJson(
|
|
|
20
20
|
* hasDensePrefillProjectionKernel?,
|
|
21
21
|
* hasQ4DecodeProjectionKernel?, hasQ4PrefillProjectionKernel?,
|
|
22
22
|
* hasAvailableQ4PrefillProjectionKernel? }
|
|
23
|
-
* @returns {{ transforms: Function[], names: string[], reason: string }}
|
|
23
|
+
* @returns {{ transforms: Function[], names: string[], reason: string, kind: string, dtypeEffect: string, evidence: string[] }}
|
|
24
24
|
*/
|
|
25
25
|
export function resolveCapabilityTransforms(capabilities, platform, graphContext) {
|
|
26
26
|
const normalizedGraphContext = graphContext ?? {};
|
|
27
27
|
const matchContext = {
|
|
28
28
|
...capabilities,
|
|
29
29
|
...normalizedGraphContext,
|
|
30
|
+
requestedActivationDtype: normalizedGraphContext.requestedActivationDtype
|
|
31
|
+
?? normalizedGraphContext.activationDtype
|
|
32
|
+
?? null,
|
|
30
33
|
activationDtype: normalizedGraphContext.activationDtype ?? null,
|
|
31
34
|
mathDtype: normalizedGraphContext.mathDtype ?? null,
|
|
32
35
|
accumDtype: normalizedGraphContext.accumDtype ?? null,
|
|
@@ -59,6 +62,9 @@ export function resolveCapabilityTransforms(capabilities, platform, graphContext
|
|
|
59
62
|
transforms,
|
|
60
63
|
names: rule.transforms,
|
|
61
64
|
reason: rule.reason,
|
|
65
|
+
kind: rule.kind,
|
|
66
|
+
dtypeEffect: rule.dtypeEffect,
|
|
67
|
+
evidence: Array.isArray(rule.evidence) ? [...rule.evidence] : [],
|
|
62
68
|
};
|
|
63
69
|
}
|
|
64
70
|
}
|
|
@@ -805,10 +805,10 @@ export function widenToF32Activations(graph, ctx) {
|
|
|
805
805
|
// When the GPU cannot compile any f16 WGSL (hasF16=false), use the full f32
|
|
806
806
|
// map that also covers f16-weight and f16-KV kernels. Otherwise use the
|
|
807
807
|
// activation-only map that preserves f16 weights/KV for precision fallback.
|
|
808
|
-
const shaderMap = ctx.capabilities?.hasF16 === false
|
|
808
|
+
const shaderMap = ctx.capabilities?.hasF16 === false || ctx.kvDtype === 'f32'
|
|
809
809
|
? FULL_F32_SHADER_MAP
|
|
810
810
|
: F16_TO_F32_ACTIVATION_MAP;
|
|
811
|
-
const fullF32 = ctx.capabilities?.hasF16 === false;
|
|
811
|
+
const fullF32 = ctx.capabilities?.hasF16 === false || ctx.kvDtype === 'f32';
|
|
812
812
|
|
|
813
813
|
const hasTargetShader = Object.values(graph.kernels).some(
|
|
814
814
|
(entry) => shaderMap.has(entry.kernel)
|
package/src/debug/config.d.ts
CHANGED
|
@@ -41,6 +41,10 @@ export interface LogHistoryEntry {
|
|
|
41
41
|
data?: unknown;
|
|
42
42
|
}
|
|
43
43
|
|
|
44
|
+
export type LogLevel = LogLevelName;
|
|
45
|
+
export type LogLevelValue = LogLevels[keyof LogLevels];
|
|
46
|
+
export type LogEntry = LogHistoryEntry;
|
|
47
|
+
|
|
44
48
|
export const LOG_LEVELS: LogLevels;
|
|
45
49
|
export const TRACE_CATEGORIES: readonly TraceCategory[];
|
|
46
50
|
|
|
@@ -1,6 +1,15 @@
|
|
|
1
1
|
export * from '../../inference/tokenizer.js';
|
|
2
2
|
export * from '../../inference/kv-cache.js';
|
|
3
|
-
export
|
|
3
|
+
export { SpeculativeDecoder } from '../../inference/speculative.js';
|
|
4
|
+
export type {
|
|
5
|
+
DraftModel,
|
|
6
|
+
MainModel,
|
|
7
|
+
SpeculativeConfig,
|
|
8
|
+
VerificationResult,
|
|
9
|
+
StepResult,
|
|
10
|
+
DecodingStats,
|
|
11
|
+
StatsWithSpeedup,
|
|
12
|
+
} from '../../inference/speculative.js';
|
|
4
13
|
export * from '../../inference/expert-router.js';
|
|
5
14
|
export * from '../../inference/moe-router.js';
|
|
6
15
|
export * from '../../inference/multi-model-network.js';
|
|
@@ -1092,7 +1092,7 @@ async function tryActivateAdapterPayload(payload) {
|
|
|
1092
1092
|
reason: 'no_adapter_payload',
|
|
1093
1093
|
};
|
|
1094
1094
|
}
|
|
1095
|
-
const { activateLoRAFromTrainingOutput } = await import('
|
|
1095
|
+
const { activateLoRAFromTrainingOutput } = await import('../../client/runtime/model-manager.js');
|
|
1096
1096
|
try {
|
|
1097
1097
|
return await activateLoRAFromTrainingOutput(payload);
|
|
1098
1098
|
} catch (error) {
|
|
@@ -1,8 +1,6 @@
|
|
|
1
1
|
import { computeNamespacedCanonicalSha256, hashBytesSha256 } from '../../utils/canonical-hash.js';
|
|
2
2
|
import { buildMerkleTreeFromLeafHashes, DEFAULT_MERKLE_BLOCK_SIZE } from './merkle.js';
|
|
3
3
|
|
|
4
|
-
export { DEFAULT_MERKLE_BLOCK_SIZE } from './merkle.js';
|
|
5
|
-
|
|
6
4
|
function asPositiveInteger(value, label) {
|
|
7
5
|
const parsed = Number(value);
|
|
8
6
|
if (!Number.isInteger(parsed) || parsed <= 0) {
|
|
@@ -20,7 +20,7 @@ import type {
|
|
|
20
20
|
ManifestWeightsRefSchema,
|
|
21
21
|
KernelPathRef,
|
|
22
22
|
ManifestInferenceSchema,
|
|
23
|
-
|
|
23
|
+
TensorRole as SchemaTensorRole,
|
|
24
24
|
} from '../../config/schema/index.js';
|
|
25
25
|
import type { FunctionalDescriptorManifest } from './functional-descriptor.js';
|
|
26
26
|
|
|
@@ -45,6 +45,8 @@ export interface AttentionOptions extends OutputBufferOptions {
|
|
|
45
45
|
indirectBuffer?: GPUBuffer | null;
|
|
46
46
|
/** Byte offset into indirect dispatch buffer (default: 0). */
|
|
47
47
|
indirectOffset?: number;
|
|
48
|
+
/** Optional f32 attention output gate tensor for profile-enabled decode epilogue fusion. */
|
|
49
|
+
outputGate?: Tensor | null;
|
|
48
50
|
}
|
|
49
51
|
|
|
50
52
|
export interface TieredAttentionOptions extends OutputBufferOptions {
|