doppler-gpu 0.2.0 → 0.2.2
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 +29 -1
- package/README.md +97 -29
- package/package.json +40 -27
- package/src/cli/config/doppler-quickstart.json +10 -0
- package/src/cli/doppler-cli.js +66 -22
- package/src/cli/doppler-quickstart.d.ts +61 -0
- package/src/cli/doppler-quickstart.js +299 -0
- package/src/cli/doppler-serve.js +395 -0
- package/src/client/doppler-api.browser.js +13 -377
- package/src/client/doppler-api.d.ts +35 -56
- package/src/client/doppler-api.js +12 -363
- package/src/client/doppler-registry.json +55 -2
- package/src/client/runtime/index.d.ts +63 -0
- package/src/client/runtime/index.js +173 -0
- package/src/client/runtime/lora.d.ts +36 -0
- package/src/client/runtime/lora.js +165 -0
- package/src/client/{doppler-provider → runtime}/model-manager.d.ts +1 -1
- package/src/client/{doppler-provider → runtime}/model-manager.js +56 -30
- package/src/client/runtime/model-session.d.ts +72 -0
- package/src/client/runtime/model-session.js +122 -0
- package/src/client/runtime/model-source.d.ts +43 -0
- package/src/client/runtime/model-source.js +101 -0
- package/src/client/{doppler-provider → runtime}/source-runtime.d.ts +8 -5
- package/src/client/runtime/source-runtime.js +919 -0
- package/src/client/runtime/storage.d.ts +5 -0
- package/src/client/runtime/storage.js +51 -0
- package/src/client/{doppler-provider → runtime}/types.d.ts +14 -2
- package/src/config/README.md +2 -0
- package/src/config/conversion/README.md +47 -7
- package/src/config/conversion/embeddinggemma/google-embeddinggemma-300m-q4k-ehf16-af32.json +67 -5
- package/src/config/conversion/gemma3/gemma-3-1b-it-f16-af32.json +62 -11
- package/src/config/conversion/gemma3/gemma-3-1b-it-q4k-ehf16-af32.json +79 -18
- package/src/config/conversion/gemma3/gemma-3-270m-it-q4k-ehf16-af32.json +61 -13
- package/src/config/conversion/gemma3/translategemma-4b-1b-enes-q4k-ehf16-af32.json +53 -10
- package/src/config/conversion/gemma3/translategemma-4b-it-q4k-ehf16-af32.json +66 -18
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32.json +256 -0
- package/src/config/conversion/gemma4/gemma-4-moe-q4k-ehf16-af32.json +49 -8
- package/src/config/conversion/gpt-oss-20b-f16-xmxfp4.json +45 -6
- package/src/config/conversion/janus/janus-pro-1b-text-q4k-ehaf16.json +45 -6
- package/src/config/conversion/lfm2/lfm2.5-1.2b-instruct-q4k-ehf16-af32.json +57 -9
- package/src/config/conversion/qwen3/qwen-3-5-0-8b-q4k-ehaf16.json +105 -26
- package/src/config/conversion/qwen3/qwen-3-5-2b-q4k-ehaf16.json +83 -33
- package/src/config/conversion/sana/sana-sprint-0.6b-f16.json +33 -5
- package/src/config/execution-contract-check.d.ts +2 -1
- package/src/config/execution-contract-check.js +187 -53
- package/src/config/kernel-path-loader.d.ts +36 -0
- package/src/config/kernel-path-loader.js +97 -10
- package/src/config/kernels/kernel-ref-digests.js +40 -24
- package/src/config/kernels/registry.js +0 -1
- package/src/config/kernels/registry.json +772 -66
- package/src/config/merge-contract-check.js +3 -0
- package/src/config/merge.d.ts +18 -1
- package/src/config/merge.js +56 -0
- package/src/config/param-categories.js +5 -0
- package/src/config/param-validator.js +42 -26
- package/src/config/platforms/apple-m3.json +1 -1
- package/src/config/platforms/loader.js +21 -16
- package/src/config/required-inference-fields-contract-check.js +16 -0
- package/src/config/runtime/compute/f16-batched.json +2 -0
- package/src/config/runtime/default.json +1 -0
- package/src/config/runtime/experiments/bench/gemma3-bench-q4k-alt.json +57 -0
- package/src/config/runtime/experiments/bench/gemma3-bench-q4k-batched.json +57 -0
- package/src/config/runtime/experiments/bench/gemma3-bench-q4k.json +2 -0
- package/src/config/runtime/experiments/bench/qwen3-bench-q4k-alt.json +57 -0
- package/src/config/runtime/experiments/bench/qwen3-bench-q4k-batched.json +57 -0
- package/src/config/runtime/experiments/debug/gemma3-browser-diagnose-q4k.json +81 -0
- package/src/config/runtime/experiments/debug/gemma3-debug-q4k.json +5 -0
- package/src/config/runtime/experiments/debug/qwen35-browser-diagnose-q4k.json +81 -0
- package/src/config/runtime/experiments/verify/gemma4-e2b-verify.json +38 -0
- package/src/config/runtime/profiles/gemma4-e2b-low-memory.json +63 -0
- package/src/config/runtime/profiles/gemma4-e2b-throughput.json +39 -0
- package/src/config/runtime/profiles/low-memory.json +1 -0
- package/src/config/runtime/profiles/turboquant-contiguous-prod.json +29 -0
- package/src/config/runtime/profiles/turboquant-contiguous.json +29 -0
- package/src/config/runtime/profiles/turboquant.json +36 -0
- package/src/config/runtime/profiles/vector-stability.json +1 -0
- package/src/config/runtime/tiers/gemma4-128gb.json +79 -0
- package/src/config/runtime/tiers/gemma4-16gb.json +15 -2
- package/src/config/runtime/tiers/gemma4-24gb.json +15 -2
- package/src/config/runtime/tiers/gemma4-32gb.json +15 -2
- package/src/config/runtime.js +11 -4
- package/src/config/schema/converter.schema.d.ts +4 -2
- package/src/config/schema/converter.schema.js +7 -5
- package/src/config/schema/doppler.schema.d.ts +26 -2
- package/src/config/schema/doppler.schema.js +77 -28
- package/src/config/schema/execution-v1.schema.d.ts +78 -5
- package/src/config/schema/execution-v1.schema.js +62 -1
- package/src/config/schema/index.d.ts +22 -29
- package/src/config/schema/index.js +15 -17
- package/src/config/schema/inference.schema.d.ts +26 -4
- package/src/config/schema/kernel-path.schema.d.ts +30 -3
- package/src/config/schema/kernel-registry.schema.d.ts +1 -1
- package/src/config/schema/kernel-registry.schema.js +1 -0
- package/src/config/schema/kvcache.schema.js +7 -0
- package/src/config/schema/loading.schema.d.ts +28 -0
- package/src/config/schema/loading.schema.js +9 -0
- package/src/config/schema/manifest.schema.d.ts +79 -5
- package/src/config/schema/manifest.schema.js +7 -1
- package/src/config/schema/platform.schema.d.ts +10 -8
- package/src/config/schema/speculation-self.schema.js +0 -5
- package/src/config/schema/support-tiers.schema.d.ts +53 -0
- package/src/config/schema/support-tiers.schema.js +128 -0
- package/src/config/source-packages/litert/gemma-4-e2b-it.json +622 -0
- package/src/config/source-packages/registry.json +15 -0
- package/src/config/support-tiers/subsystems.json +502 -0
- package/src/config/transforms/capability-transform-resolver.d.ts +16 -5
- package/src/config/transforms/capability-transform-resolver.js +44 -9
- package/src/config/transforms/execution-graph-transforms.d.ts +113 -2
- package/src/config/transforms/execution-graph-transforms.js +1220 -8
- package/src/converter/conversion-plan.js +20 -8
- package/src/converter/core.d.ts +48 -1
- package/src/converter/core.js +249 -62
- package/src/converter/parsers/gguf.d.ts +10 -1
- package/src/converter/rope-config.d.ts +1 -0
- package/src/converter/rope-config.js +75 -5
- package/src/converter/shard-packer.js +2 -2
- package/src/debug/tensor.js +13 -1
- package/src/{adapters → experimental/adapters}/adapter-manager.d.ts +1 -1
- package/src/{adapters → experimental/adapters}/adapter-manager.js +2 -2
- package/src/{adapters → experimental/adapters}/adapter-manifest.d.ts +1 -1
- package/src/{adapters → experimental/adapters}/adapter-manifest.js +1 -1
- package/src/{adapters → experimental/adapters}/adapter-registry.d.ts +1 -1
- package/src/{adapters → experimental/adapters}/adapter-registry.js +1 -1
- package/src/{adapters → experimental/adapters}/lora-loader.d.ts +1 -1
- package/src/{adapters → experimental/adapters}/lora-loader.js +2 -2
- package/src/{bridge → experimental/bridge}/extension-client.js +2 -2
- package/src/{bridge → experimental/bridge}/node-bridge.js +2 -4
- package/src/{bridge → experimental/bridge}/protocol.js +1 -1
- package/src/{browser → experimental/browser}/browser-converter.d.ts +4 -4
- package/src/{browser → experimental/browser}/browser-converter.js +12 -12
- package/src/{browser → experimental/browser}/gguf-importer.d.ts +1 -1
- package/src/{browser → experimental/browser}/gguf-importer.js +8 -8
- package/src/{browser → experimental/browser}/gguf-parser-browser.d.ts +2 -2
- package/src/{browser → experimental/browser}/gguf-parser-browser.js +3 -3
- package/src/{browser → experimental/browser}/quantization.d.ts +3 -3
- package/src/{browser → experimental/browser}/quantization.js +4 -4
- package/src/{browser → experimental/browser}/safetensors-parser-browser.d.ts +3 -3
- package/src/{browser → experimental/browser}/safetensors-parser-browser.js +3 -3
- package/src/{browser → experimental/browser}/shard-io-browser.d.ts +1 -1
- package/src/{browser → experimental/browser}/shard-io-browser.js +2 -2
- package/src/{browser → experimental/browser}/tensor-source-download.js +3 -3
- package/src/experimental/diffusion/index.d.ts +8 -0
- package/src/experimental/diffusion/index.js +8 -0
- package/src/{distribution → experimental/distribution}/p2p-transport-contract.js +1 -1
- package/src/{distribution → experimental/distribution}/p2p-transport.js +1 -1
- package/src/{distribution → experimental/distribution}/shard-delivery.js +5 -5
- package/src/experimental/energy/index.d.ts +2 -0
- package/src/experimental/energy/index.js +2 -0
- package/src/{hotswap → experimental/hotswap}/intent-bundle.d.ts +1 -1
- package/src/{hotswap → experimental/hotswap}/intent-bundle.js +2 -2
- package/src/{hotswap → experimental/hotswap}/manifest.d.ts +2 -2
- package/src/{hotswap → experimental/hotswap}/manifest.js +1 -1
- package/src/{hotswap → experimental/hotswap}/model-swap.js +1 -1
- package/src/{hotswap → experimental/hotswap}/runtime.d.ts +1 -1
- package/src/experimental/orchestration/index.d.ts +12 -0
- package/src/experimental/orchestration/index.js +12 -0
- package/src/{training → experimental/training}/README.md +1 -1
- package/src/{training → experimental/training}/artifacts.d.ts +3 -3
- package/src/{training → experimental/training}/artifacts.js +3 -3
- package/src/{training → experimental/training}/attention-backward.d.ts +1 -1
- package/src/{training → experimental/training}/attention-backward.js +3 -3
- package/src/{training → experimental/training}/attention-forward.d.ts +2 -2
- package/src/{training → experimental/training}/attention-forward.js +1 -1
- package/src/{training → experimental/training}/autograd.d.ts +2 -2
- package/src/{training → experimental/training}/autograd.js +7 -7
- package/src/{training → experimental/training}/backward-kernels.js +1 -1
- package/src/{training → experimental/training}/checkpoint.js +1 -1
- package/src/{training → experimental/training}/clip.d.ts +2 -2
- package/src/{training → experimental/training}/clip.js +3 -3
- package/src/{training → experimental/training}/export.d.ts +2 -2
- package/src/{training → experimental/training}/export.js +2 -2
- package/src/{training → experimental/training}/lora-pipeline.js +4 -4
- package/src/{training → experimental/training}/lora.d.ts +1 -1
- package/src/{training → experimental/training}/lora.js +4 -4
- package/src/{training → experimental/training}/loss-scaling.d.ts +2 -2
- package/src/{training → experimental/training}/loss-scaling.js +3 -3
- package/src/{training → experimental/training}/loss.d.ts +2 -2
- package/src/{training → experimental/training}/loss.js +1 -1
- package/src/{training → experimental/training}/operator-artifacts.js +1 -1
- package/src/{training → experimental/training}/optimizer.d.ts +2 -2
- package/src/{training → experimental/training}/optimizer.js +3 -3
- package/src/{training → experimental/training}/runner.d.ts +2 -2
- package/src/{training → experimental/training}/runner.js +5 -5
- package/src/{training → experimental/training}/suite.js +16 -16
- package/src/{training → experimental/training}/tensor-factory.d.ts +1 -1
- package/src/{training → experimental/training}/tensor-factory.js +2 -2
- package/src/{training → experimental/training}/trainer.d.ts +3 -3
- package/src/{training → experimental/training}/trainer.js +3 -3
- package/src/{training → experimental/training}/ul_dataset.d.ts +2 -2
- package/src/{training → experimental/training}/ul_dataset.js +1 -1
- package/src/{training → experimental/training}/ul_schedule.d.ts +1 -1
- package/src/{training → experimental/training}/workloads.js +4 -4
- package/src/formats/index.d.ts +24 -0
- package/src/formats/index.js +6 -0
- package/src/formats/litert/index.d.ts +1 -0
- package/src/formats/litert/index.js +1 -0
- package/src/formats/litert/types.d.ts +87 -0
- package/src/formats/litert/types.js +568 -0
- package/src/formats/rdrr/classification.js +16 -8
- package/src/formats/rdrr/parsing.d.ts +3 -2
- package/src/formats/rdrr/types.d.ts +33 -0
- package/src/formats/rdrr/validation.js +16 -5
- package/src/formats/tflite/index.d.ts +1 -0
- package/src/formats/tflite/index.js +1 -0
- package/src/formats/tflite/types.d.ts +84 -0
- package/src/formats/tflite/types.js +633 -0
- package/src/generation/index.d.ts +16 -14
- package/src/generation/index.js +1 -32
- package/src/gpu/command-recorder.d.ts +17 -1
- package/src/gpu/command-recorder.js +47 -10
- package/src/gpu/device.d.ts +3 -1
- package/src/gpu/device.js +17 -11
- package/src/gpu/kernel-selector.js +3 -1
- package/src/gpu/kernels/attention.d.ts +2 -0
- package/src/gpu/kernels/attention.js +359 -14
- package/src/gpu/kernels/attention.wgsl +20 -1
- package/src/gpu/kernels/attention_decode_contiguous_turboquant_f16kv.wgsl +180 -0
- package/src/gpu/kernels/attention_decode_contiguous_turboquant_prod_f16kv.wgsl +192 -0
- package/src/gpu/kernels/attention_decode_online_f16.wgsl +36 -18
- package/src/gpu/kernels/attention_decode_online_f16kv.wgsl +36 -18
- package/src/gpu/kernels/attention_decode_tiered_turboquant_f16kv.wgsl +269 -0
- package/src/gpu/kernels/attention_decode_tiered_turboquant_prod_f16kv.wgsl +281 -0
- package/src/gpu/kernels/attention_head256_f16kv.wgsl +269 -0
- package/src/gpu/kernels/attention_streaming.wgsl +22 -3
- package/src/gpu/kernels/attention_streaming_f16.wgsl +1 -1
- package/src/gpu/kernels/attention_streaming_f16kv.wgsl +1 -1
- package/src/gpu/kernels/backward/attention_backward.d.ts +14 -1
- package/src/gpu/kernels/check-hot-vocab-stop.d.ts +18 -0
- package/src/gpu/kernels/check-hot-vocab-stop.js +108 -0
- package/src/gpu/kernels/codegen/patches/attention_streaming_f16.from.attention_streaming.diff +1 -1
- package/src/gpu/kernels/codegen/patches/attention_streaming_f16kv.from.attention_streaming.diff +1 -1
- package/src/gpu/kernels/codegen/patches/rope_f16.from.rope.diff +126 -194
- package/src/gpu/kernels/codegen/wgsl-variants.js +24 -6
- package/src/gpu/kernels/depthwise_conv1d.d.ts +23 -0
- package/src/gpu/kernels/depthwise_conv1d.js +51 -0
- package/src/gpu/kernels/depthwise_conv1d.wgsl +46 -0
- package/src/gpu/kernels/feature-check.js +0 -1
- package/src/gpu/kernels/fused_ffn.js +67 -22
- package/src/gpu/kernels/fused_ffn_q4k.wgsl +48 -43
- package/src/gpu/kernels/fused_ffn_q4k_f16.wgsl +329 -0
- package/src/gpu/kernels/fused_matmul_q4.wgsl +417 -0
- package/src/gpu/kernels/fused_matmul_q4_batched_multicol_shared.wgsl +174 -0
- package/src/gpu/kernels/gather.d.ts +4 -0
- package/src/gpu/kernels/gather.js +13 -2
- package/src/gpu/kernels/gather.wgsl +5 -3
- package/src/gpu/kernels/gather_f16.wgsl +5 -3
- package/src/gpu/kernels/index.d.ts +7 -0
- package/src/gpu/kernels/index.js +6 -0
- package/src/gpu/kernels/kernel-configs.js +1 -0
- package/src/gpu/kernels/kv-quantize.js +87 -55
- package/src/gpu/kernels/kv_quantize_turboquant.wgsl +153 -0
- package/src/gpu/kernels/kv_quantize_turboquant_prod.wgsl +230 -0
- package/src/gpu/kernels/linear-attention-core.d.ts +3 -0
- package/src/gpu/kernels/linear-attention-core.js +145 -56
- package/src/gpu/kernels/matmul-dispatch.js +3 -0
- package/src/gpu/kernels/matmul-selection.js +31 -42
- package/src/gpu/kernels/matmul.d.ts +3 -0
- package/src/gpu/kernels/matmul.js +72 -34
- package/src/gpu/kernels/residual.d.ts +2 -0
- package/src/gpu/kernels/residual.js +43 -4
- package/src/gpu/kernels/rope.d.ts +2 -0
- package/src/gpu/kernels/rope.js +19 -6
- package/src/gpu/kernels/rope.wgsl +24 -12
- package/src/gpu/kernels/rope_f16.wgsl +97 -38
- package/src/gpu/kernels/scale.d.ts +5 -0
- package/src/gpu/kernels/scale.js +6 -1
- package/src/gpu/kernels/scale.wgsl +3 -1
- package/src/gpu/kernels/scale_f16.wgsl +32 -0
- package/src/gpu/kernels/turboquant-codebook.js +564 -0
- package/src/gpu/kernels/uniform-utils.js +26 -0
- package/src/gpu/kernels/utils.js +6 -1
- package/src/gpu/submit-probe.d.ts +12 -0
- package/src/gpu/submit-probe.js +77 -0
- package/src/gpu/weight-buffer.d.ts +26 -3
- package/src/gpu/weight-buffer.js +62 -2
- package/src/index-browser.d.ts +1 -91
- package/src/index-browser.js +0 -66
- package/src/index.d.ts +1 -102
- package/src/index.js +0 -74
- package/src/inference/browser-harness-model-helpers.d.ts +3 -0
- package/src/inference/browser-harness-model-helpers.js +134 -11
- package/src/inference/browser-harness-runtime-helpers.js +1 -9
- package/src/inference/browser-harness-text-helpers.js +441 -70
- package/src/inference/browser-harness.js +61 -4
- package/src/inference/decode-ring.d.ts +2 -0
- package/src/inference/decode-ring.js +20 -0
- package/src/inference/kv-cache/index.d.ts +3 -0
- package/src/inference/kv-cache/index.js +2 -0
- package/src/inference/kv-cache/mixed-geometry.d.ts +84 -0
- package/src/inference/kv-cache/mixed-geometry.js +422 -0
- package/src/inference/kv-cache/quantized.d.ts +56 -0
- package/src/inference/kv-cache/quantized.js +383 -0
- package/src/inference/kv-cache/tiered.d.ts +20 -3
- package/src/inference/kv-cache/tiered.js +203 -65
- package/src/inference/kv-cache/types.d.ts +49 -6
- package/src/inference/kv-cache.d.ts +3 -0
- package/src/inference/kv-cache.js +2 -0
- package/src/inference/multi-model-network.js +2 -1
- package/src/inference/pipelines/audio/gemma4.js +655 -0
- package/src/inference/pipelines/audio/index.js +39 -0
- package/src/inference/pipelines/audio/mel.js +171 -0
- package/src/inference/pipelines/context.js +9 -1
- package/src/inference/pipelines/diffusion/text-encoder-gpu.js +2 -0
- package/src/inference/pipelines/shared/clipped-linear.js +46 -0
- package/src/inference/pipelines/text/attention/dispatch-params.d.ts +2 -1
- package/src/inference/pipelines/text/attention/dispatch-params.js +52 -8
- package/src/inference/pipelines/text/attention/output-projection.js +3 -3
- package/src/inference/pipelines/text/attention/precision-contract.d.ts +36 -0
- package/src/inference/pipelines/text/attention/precision-contract.js +115 -0
- package/src/inference/pipelines/text/attention/projections.d.ts +23 -0
- package/src/inference/pipelines/text/attention/projections.js +182 -15
- package/src/inference/pipelines/text/attention/quant-options.js +54 -0
- package/src/inference/pipelines/text/attention/record.js +257 -69
- package/src/inference/pipelines/text/attention/run.js +306 -94
- package/src/inference/pipelines/text/attention/types.d.ts +11 -0
- package/src/inference/pipelines/text/chat-format.d.ts +8 -1
- package/src/inference/pipelines/text/chat-format.js +89 -5
- package/src/inference/pipelines/text/config.d.ts +78 -26
- package/src/inference/pipelines/text/config.js +722 -28
- package/src/inference/pipelines/text/dtype-contract.d.ts +15 -0
- package/src/inference/pipelines/text/dtype-contract.js +38 -0
- package/src/inference/pipelines/text/embed.d.ts +4 -0
- package/src/inference/pipelines/text/embed.js +165 -172
- package/src/inference/pipelines/text/execution-plan.d.ts +19 -0
- package/src/inference/pipelines/text/execution-plan.js +112 -20
- package/src/inference/pipelines/text/execution-runtime-builders.d.ts +8 -4
- package/src/inference/pipelines/text/execution-runtime-builders.js +90 -49
- package/src/inference/pipelines/text/execution-v1.d.ts +23 -4
- package/src/inference/pipelines/text/execution-v1.js +411 -33
- package/src/inference/pipelines/text/ffn/dense.d.ts +59 -0
- package/src/inference/pipelines/text/ffn/dense.js +575 -63
- package/src/inference/pipelines/text/ffn/moe.js +1 -0
- package/src/inference/pipelines/text/ffn/sandwich.js +20 -4
- package/src/inference/pipelines/text/ffn/standard.js +28 -4
- package/src/inference/pipelines/text/finiteness-policy.d.ts +4 -0
- package/src/inference/pipelines/text/finiteness-policy.js +14 -11
- package/src/inference/pipelines/text/generator-helpers.d.ts +7 -0
- package/src/inference/pipelines/text/generator-helpers.js +39 -4
- package/src/inference/pipelines/text/generator-runtime.js +45 -8
- package/src/inference/pipelines/text/generator-steps.js +407 -75
- package/src/inference/pipelines/text/generator.d.ts +6 -0
- package/src/inference/pipelines/text/generator.js +1179 -305
- package/src/inference/pipelines/text/init.d.ts +39 -7
- package/src/inference/pipelines/text/init.js +525 -110
- package/src/inference/pipelines/text/layer-plan.d.ts +1 -0
- package/src/inference/pipelines/text/layer-plan.js +2 -0
- package/src/inference/pipelines/text/layer.js +402 -19
- package/src/inference/pipelines/text/linear-attention.d.ts +4 -0
- package/src/inference/pipelines/text/linear-attention.js +112 -41
- package/src/inference/pipelines/text/logits/gpu.d.ts +2 -1
- package/src/inference/pipelines/text/logits/gpu.js +334 -15
- package/src/inference/pipelines/text/logits/index.d.ts +11 -0
- package/src/inference/pipelines/text/logits/index.js +227 -29
- package/src/inference/pipelines/text/logits/types.d.ts +3 -0
- package/src/inference/pipelines/text/logits/utils.js +3 -1
- package/src/inference/pipelines/text/lora-apply.js +3 -3
- package/src/inference/pipelines/text/model-load.d.ts +2 -1
- package/src/inference/pipelines/text/model-load.js +205 -178
- package/src/inference/pipelines/text/moe-gpu.js +8 -0
- package/src/inference/pipelines/text/moe-impl.d.ts +2 -0
- package/src/inference/pipelines/text/ops.d.ts +16 -2
- package/src/inference/pipelines/text/ops.js +11 -2
- package/src/inference/pipelines/text/per-layer-inputs.d.ts +107 -0
- package/src/inference/pipelines/text/per-layer-inputs.js +1581 -0
- package/src/inference/pipelines/text/probes.js +6 -0
- package/src/inference/pipelines/text/sampling-config.js +9 -1
- package/src/inference/pipelines/text/stage-names.js +3 -0
- package/src/inference/pipelines/text/state.d.ts +12 -3
- package/src/inference/pipelines/text/state.js +19 -0
- package/src/inference/pipelines/text/types.d.ts +79 -7
- package/src/inference/pipelines/text/weights.d.ts +2 -2
- package/src/inference/pipelines/text/weights.js +4 -0
- package/src/inference/pipelines/text.d.ts +6 -0
- package/src/inference/pipelines/text.js +860 -11
- package/src/inference/pipelines/video/frame-sampler.js +36 -0
- package/src/inference/pipelines/video/index.js +87 -0
- package/src/inference/pipelines/vision/encoder.js +27 -6
- package/src/inference/pipelines/vision/gemma4.js +731 -0
- package/src/inference/pipelines/vision/image-preprocess.js +22 -7
- package/src/inference/pipelines/vision/index.js +35 -10
- package/src/inference/pipelines/vision/ops.js +31 -26
- package/src/inference/pipelines/vision/patch-embed.js +22 -10
- package/src/inference/runtime-model.d.ts +89 -0
- package/src/inference/runtime-model.js +152 -0
- package/src/inference/test-harness.d.ts +8 -1
- package/src/inference/test-harness.js +54 -20
- package/src/inference/tokenizer.d.ts +5 -0
- package/src/inference/tokenizer.js +43 -0
- package/src/inference/tokenizers/base.d.ts +5 -0
- package/src/inference/tokenizers/base.js +5 -0
- package/src/inference/tokenizers/behavior-flags.d.ts +9 -0
- package/src/inference/tokenizers/behavior-flags.js +100 -0
- package/src/inference/tokenizers/bundled.d.ts +4 -0
- package/src/inference/tokenizers/bundled.js +110 -0
- package/src/inference/tokenizers/sentencepiece.d.ts +2 -0
- package/src/inference/tokenizers/sentencepiece.js +16 -0
- package/src/inference/tokenizers/types.d.ts +2 -0
- package/src/loader/doppler-loader.d.ts +10 -3
- package/src/loader/doppler-loader.js +103 -5
- package/src/loader/embedding-loader.d.ts +5 -1
- package/src/loader/embedding-loader.js +57 -6
- package/src/loader/final-weights-loader.js +44 -5
- package/src/loader/layer-loader.js +210 -8
- package/src/loader/loader-state.js +2 -2
- package/src/loader/loader-types.d.ts +17 -1
- package/src/loader/memory-monitor.d.ts +8 -0
- package/src/loader/memory-monitor.js +21 -0
- package/src/loader/multi-model-loader.d.ts +9 -1
- package/src/loader/multi-model-loader.js +9 -1
- package/src/loader/per-layer-input-loader.d.ts +30 -0
- package/src/loader/per-layer-input-loader.js +478 -0
- package/src/loader/shard-cache.d.ts +4 -3
- package/src/loader/shard-resolver.js +2 -0
- package/src/loader/tensors/source-transform.d.ts +50 -0
- package/src/loader/tensors/source-transform.js +776 -0
- package/src/loader/tensors/tensor-loader.d.ts +12 -0
- package/src/loader/tensors/tensor-loader.js +208 -123
- package/src/loader/tensors/tensor-reader.d.ts +8 -0
- package/src/loader/tensors/tensor-reader.js +283 -15
- package/src/loaders/index.d.ts +4 -0
- package/src/loaders/index.js +4 -0
- package/src/memory/buffer-pool.d.ts +4 -0
- package/src/memory/buffer-pool.js +36 -0
- package/src/rules/execution-rules-contract-check.d.ts +4 -0
- package/src/rules/execution-rules-contract-check.js +150 -24
- package/src/rules/inference/attention.rules.json +15 -0
- package/src/rules/inference/capability-transforms.rules.json +30 -0
- package/src/rules/inference/execution.rules.json +54 -2
- package/src/rules/inference/ffn.rules.json +33 -0
- package/src/rules/kernels/attention.rules.json +7 -0
- package/src/rules/kernels/fused-ffn.rules.json +4 -2
- package/src/rules/kernels/kv_quantize.rules.json +2 -0
- package/src/rules/kernels/matmul.rules.json +14 -0
- package/src/rules/kernels/scale.rules.json +2 -0
- package/src/rules/loader/tensor-loader.rules.json +1 -0
- package/src/storage/artifact-storage-context.d.ts +53 -0
- package/src/storage/artifact-storage-context.js +560 -0
- package/src/storage/downloader.js +7 -1
- package/src/storage/index.d.ts +18 -0
- package/src/storage/index.js +13 -0
- package/src/storage/preflight.js +22 -0
- package/src/storage/quickstart-downloader.js +35 -2
- package/src/storage/registry.d.ts +1 -0
- package/src/storage/registry.js +26 -4
- package/src/storage/shard-manager.d.ts +4 -0
- package/src/storage/shard-manager.js +25 -6
- package/src/storage/source-artifact-store.d.ts +9 -0
- package/src/storage/source-artifact-store.js +134 -0
- package/src/tooling/browser-command-runner.js +12 -5
- package/src/tooling/command-api-family-normalizers.js +144 -0
- package/src/tooling/command-api-helpers.js +1 -0
- package/src/tooling/command-api.d.ts +32 -9
- package/src/tooling/command-api.js +0 -25
- package/src/tooling/command-runner-shared.js +1 -2
- package/src/tooling/command-runner.html +5 -2
- package/src/tooling/diagnose-runner.js +1 -1
- package/src/tooling/kernel-path-builder/index.js +48 -32
- package/src/tooling/lean-execution-contract.d.ts +1 -1
- package/src/tooling/lean-execution-contract.js +2 -149
- package/src/tooling/litert-package-runtime.d.ts +46 -0
- package/src/tooling/litert-package-runtime.js +950 -0
- package/src/tooling/node-browser-command-runner.d.ts +13 -0
- package/src/tooling/node-browser-command-runner.js +163 -3
- package/src/tooling/node-command-runner.js +14 -5
- package/src/tooling/node-converter.js +226 -63
- package/src/tooling/node-source-runtime.d.ts +8 -3
- package/src/tooling/node-source-runtime.js +561 -236
- package/src/tooling/node-webgpu.js +7 -30
- package/src/tooling/opfs-cache.d.ts +18 -1
- package/src/tooling/opfs-cache.js +126 -9
- package/src/tooling/runtime-input-composition.d.ts +0 -1
- package/src/tooling/runtime-input-composition.js +0 -5
- package/src/tooling/source-artifact-adapter.d.ts +134 -0
- package/src/tooling/source-artifact-adapter.js +366 -0
- package/src/tooling/source-package-profiles.d.ts +33 -0
- package/src/tooling/source-package-profiles.js +57 -0
- package/src/tooling/source-runtime-bundle.d.ts +46 -6
- package/src/tooling/source-runtime-bundle.js +371 -60
- package/src/tooling/source-runtime-converter-config.d.ts +24 -0
- package/src/tooling/source-runtime-converter-config.js +297 -0
- package/src/tooling-experimental-exports.browser.js +2 -0
- package/src/tooling-experimental-exports.d.ts +7 -0
- package/src/tooling-experimental-exports.js +7 -0
- package/src/tooling-experimental-exports.shared.d.ts +26 -0
- package/src/tooling-experimental-exports.shared.js +29 -0
- package/src/tooling-exports.shared.d.ts +0 -23
- package/src/tooling-exports.shared.js +6 -28
- package/src/version.js +1 -1
- package/src/client/doppler-provider/generation.d.ts +0 -25
- package/src/client/doppler-provider/generation.js +0 -126
- package/src/client/doppler-provider/index.d.ts +0 -2
- package/src/client/doppler-provider/index.js +0 -3
- package/src/client/doppler-provider/provider.d.ts +0 -5
- package/src/client/doppler-provider/provider.js +0 -102
- package/src/client/doppler-provider/source-runtime.js +0 -677
- package/src/client/doppler-provider.d.ts +0 -46
- package/src/client/doppler-provider.js +0 -36
- package/src/config/schema/inference-defaults.schema.d.ts +0 -276
- package/src/config/schema/inference-defaults.schema.js +0 -189
- package/src/diffusion/index.d.ts +0 -8
- package/src/diffusion/index.js +0 -8
- package/src/energy/index.d.ts +0 -2
- package/src/energy/index.js +0 -2
- package/src/index-internal.d.ts +0 -2
- package/src/index-internal.js +0 -2
- /package/src/client/{doppler-provider → runtime}/types.js +0 -0
- /package/src/{adapters → experimental/adapters}/index.d.ts +0 -0
- /package/src/{adapters → experimental/adapters}/index.js +0 -0
- /package/src/{bridge → experimental/bridge}/extension/background.d.ts +0 -0
- /package/src/{bridge → experimental/bridge}/extension/background.js +0 -0
- /package/src/{bridge → experimental/bridge}/extension/manifest.json +0 -0
- /package/src/{bridge → experimental/bridge}/extension-client.d.ts +0 -0
- /package/src/{bridge → experimental/bridge}/index.d.ts +0 -0
- /package/src/{bridge → experimental/bridge}/index.js +0 -0
- /package/src/{bridge → experimental/bridge}/protocol.d.ts +0 -0
- /package/src/{browser → experimental/browser}/file-picker.d.ts +0 -0
- /package/src/{browser → experimental/browser}/file-picker.js +0 -0
- /package/src/{browser → experimental/browser}/tensor-source-download.d.ts +0 -0
- /package/src/{browser → experimental/browser}/tensor-source-file.d.ts +0 -0
- /package/src/{browser → experimental/browser}/tensor-source-file.js +0 -0
- /package/src/{browser → experimental/browser}/tensor-source-http.d.ts +0 -0
- /package/src/{browser → experimental/browser}/tensor-source-http.js +0 -0
- /package/src/{diffusion → experimental/diffusion}/config/expansion-lanes.json +0 -0
- /package/src/{diffusion → experimental/diffusion}/image-regression.d.ts +0 -0
- /package/src/{diffusion → experimental/diffusion}/image-regression.js +0 -0
- /package/src/{distribution → experimental/distribution}/p2p-control-plane.d.ts +0 -0
- /package/src/{distribution → experimental/distribution}/p2p-control-plane.js +0 -0
- /package/src/{distribution → experimental/distribution}/p2p-observability.d.ts +0 -0
- /package/src/{distribution → experimental/distribution}/p2p-observability.js +0 -0
- /package/src/{distribution → experimental/distribution}/p2p-transport-contract.d.ts +0 -0
- /package/src/{distribution → experimental/distribution}/p2p-webrtc-browser.d.ts +0 -0
- /package/src/{distribution → experimental/distribution}/p2p-webrtc-browser.js +0 -0
- /package/src/{distribution → experimental/distribution}/shard-delivery.d.ts +0 -0
- /package/src/{hotswap → experimental/hotswap}/runtime.js +0 -0
- /package/src/{training → experimental/training}/checkpoint-watch.d.ts +0 -0
- /package/src/{training → experimental/training}/checkpoint-watch.js +0 -0
- /package/src/{training → experimental/training}/checkpoint.d.ts +0 -0
- /package/src/{training → experimental/training}/dataloader.d.ts +0 -0
- /package/src/{training → experimental/training}/dataloader.js +0 -0
- /package/src/{training → experimental/training}/datasets/index.d.ts +0 -0
- /package/src/{training → experimental/training}/datasets/index.js +0 -0
- /package/src/{training → experimental/training}/datasets/jsonl.d.ts +0 -0
- /package/src/{training → experimental/training}/datasets/jsonl.js +0 -0
- /package/src/{training → experimental/training}/datasets/reploid.d.ts +0 -0
- /package/src/{training → experimental/training}/datasets/reploid.js +0 -0
- /package/src/{training → experimental/training}/datasets/text-pairs.d.ts +0 -0
- /package/src/{training → experimental/training}/datasets/text-pairs.js +0 -0
- /package/src/{training → experimental/training}/datasets/token-batch.d.ts +0 -0
- /package/src/{training → experimental/training}/datasets/token-batch.js +0 -0
- /package/src/{training → experimental/training}/datasets/translation-pairs.d.ts +0 -0
- /package/src/{training → experimental/training}/datasets/translation-pairs.js +0 -0
- /package/src/{training → experimental/training}/distillation/artifacts.d.ts +0 -0
- /package/src/{training → experimental/training}/distillation/artifacts.js +0 -0
- /package/src/{training → experimental/training}/distillation/checkpoint-watch.d.ts +0 -0
- /package/src/{training → experimental/training}/distillation/checkpoint-watch.js +0 -0
- /package/src/{training → experimental/training}/distillation/dataset.d.ts +0 -0
- /package/src/{training → experimental/training}/distillation/dataset.js +0 -0
- /package/src/{training → experimental/training}/distillation/eval.d.ts +0 -0
- /package/src/{training → experimental/training}/distillation/eval.js +0 -0
- /package/src/{training → experimental/training}/distillation/index.d.ts +0 -0
- /package/src/{training → experimental/training}/distillation/index.js +0 -0
- /package/src/{training → experimental/training}/distillation/runtime.d.ts +0 -0
- /package/src/{training → experimental/training}/distillation/runtime.js +0 -0
- /package/src/{training → experimental/training}/distillation/scoreboard.d.ts +0 -0
- /package/src/{training → experimental/training}/distillation/scoreboard.js +0 -0
- /package/src/{training → experimental/training}/distillation/stage-a.d.ts +0 -0
- /package/src/{training → experimental/training}/distillation/stage-a.js +0 -0
- /package/src/{training → experimental/training}/distillation/stage-b.d.ts +0 -0
- /package/src/{training → experimental/training}/distillation/stage-b.js +0 -0
- /package/src/{training → experimental/training}/distillation/student-fixture.d.ts +0 -0
- /package/src/{training → experimental/training}/distillation/student-fixture.js +0 -0
- /package/src/{training → experimental/training}/distillation/suite-data.d.ts +0 -0
- /package/src/{training → experimental/training}/distillation/suite-data.js +0 -0
- /package/src/{training → experimental/training}/index.d.ts +0 -0
- /package/src/{training → experimental/training}/index.js +0 -0
- /package/src/{training → experimental/training}/lora-pipeline.d.ts +0 -0
- /package/src/{training → experimental/training}/objectives/base.d.ts +0 -0
- /package/src/{training → experimental/training}/objectives/base.js +0 -0
- /package/src/{training → experimental/training}/objectives/cross_entropy.d.ts +0 -0
- /package/src/{training → experimental/training}/objectives/cross_entropy.js +0 -0
- /package/src/{training → experimental/training}/objectives/distill_kd.d.ts +0 -0
- /package/src/{training → experimental/training}/objectives/distill_kd.js +0 -0
- /package/src/{training → experimental/training}/objectives/distill_triplet.d.ts +0 -0
- /package/src/{training → experimental/training}/objectives/distill_triplet.js +0 -0
- /package/src/{training → experimental/training}/objectives/index.d.ts +0 -0
- /package/src/{training → experimental/training}/objectives/index.js +0 -0
- /package/src/{training → experimental/training}/objectives/ul_stage1_joint.d.ts +0 -0
- /package/src/{training → experimental/training}/objectives/ul_stage1_joint.js +0 -0
- /package/src/{training → experimental/training}/objectives/ul_stage2_base.d.ts +0 -0
- /package/src/{training → experimental/training}/objectives/ul_stage2_base.js +0 -0
- /package/src/{training → experimental/training}/operator-artifacts.d.ts +0 -0
- /package/src/{training → experimental/training}/operator-command.d.ts +0 -0
- /package/src/{training → experimental/training}/operator-command.js +0 -0
- /package/src/{training → experimental/training}/operator-eval.d.ts +0 -0
- /package/src/{training → experimental/training}/operator-eval.js +0 -0
- /package/src/{training → experimental/training}/operator-scoreboard.d.ts +0 -0
- /package/src/{training → experimental/training}/operator-scoreboard.js +0 -0
- /package/src/{training → experimental/training}/suite.d.ts +0 -0
- /package/src/{training → experimental/training}/ul_schedule.js +0 -0
- /package/src/{training → experimental/training}/workload-packs/distill-translategemma-medium.json +0 -0
- /package/src/{training → experimental/training}/workload-packs/distill-translategemma-tiny.json +0 -0
- /package/src/{training → experimental/training}/workload-packs/lora-toy-medium.json +0 -0
- /package/src/{training → experimental/training}/workload-packs/lora-toy-tiny.json +0 -0
- /package/src/{training → experimental/training}/workload-packs/registry.json +0 -0
- /package/src/{training → experimental/training}/workload-packs/ul-training-medium.json +0 -0
- /package/src/{training → experimental/training}/workload-packs/ul-training-tiny.json +0 -0
- /package/src/{training → experimental/training}/workloads.d.ts +0 -0
|
@@ -0,0 +1,655 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Gemma 4 Conformer Audio Encoder
|
|
3
|
+
*
|
|
4
|
+
* Architecture:
|
|
5
|
+
* mel spectrogram -> 2-stage conv subsampling -> linear projection
|
|
6
|
+
* -> 12 conformer layers (half-FFN + self-attn + depthwise conv1d + half-FFN + norm)
|
|
7
|
+
* -> output projection -> audio embedding projection
|
|
8
|
+
*
|
|
9
|
+
* All linears use clipped ranges (input_min/max/output_min/max).
|
|
10
|
+
* Self-attention uses relative position encoding with per_dim_scale and logit capping.
|
|
11
|
+
*/
|
|
12
|
+
|
|
13
|
+
import { log } from '../../../debug/index.js';
|
|
14
|
+
import { getDevice } from '../../../gpu/device.js';
|
|
15
|
+
import { createTensor } from '../../../gpu/tensor.js';
|
|
16
|
+
import {
|
|
17
|
+
runMatmul,
|
|
18
|
+
runRMSNorm,
|
|
19
|
+
runResidualAdd,
|
|
20
|
+
runSiLU,
|
|
21
|
+
} from '../../../gpu/kernel-selector.js';
|
|
22
|
+
import { runConv2D } from '../../../gpu/kernels/conv2d.js';
|
|
23
|
+
import { runDepthwiseConv1D } from '../../../gpu/kernels/depthwise_conv1d.js';
|
|
24
|
+
import { acquireBuffer, readBuffer, releaseBuffer, uploadData } from '../../../memory/buffer-pool.js';
|
|
25
|
+
import { runClippableLinear } from '../shared/clipped-linear.js';
|
|
26
|
+
|
|
27
|
+
function reshapeTensor(tensor, shape, label) {
|
|
28
|
+
return createTensor(tensor.buffer, tensor.dtype, shape, label ?? tensor.label);
|
|
29
|
+
}
|
|
30
|
+
|
|
31
|
+
// ---------------------------------------------------------------------------
|
|
32
|
+
// Subsampling: 2-stage Conv2D + LayerNorm + Linear projection
|
|
33
|
+
// ---------------------------------------------------------------------------
|
|
34
|
+
|
|
35
|
+
async function runSubsampling(melFeatures, numFrames, nMels, weights, audioConfig) {
|
|
36
|
+
const hiddenSize = audioConfig.hiddenSize;
|
|
37
|
+
const convChannels = audioConfig.subsamplingConvChannels;
|
|
38
|
+
|
|
39
|
+
// Reshape mel [numFrames, nMels] -> [1, numFrames, nMels] for conv2d (treat as single-channel image)
|
|
40
|
+
const inputBuffer = await uploadData(melFeatures, 'audio_mel_input');
|
|
41
|
+
let currentTensor = createTensor(inputBuffer, 'f32', [1, numFrames, nMels], 'audio_mel');
|
|
42
|
+
|
|
43
|
+
let currentHeight = numFrames;
|
|
44
|
+
let currentWidth = nMels;
|
|
45
|
+
let currentChannels = 1;
|
|
46
|
+
|
|
47
|
+
// Conv layer 0: [1, H, W] -> [convChannels[0], H/2, W/2] with stride=2, kernel=3, pad=1
|
|
48
|
+
const conv0Result = await runConv2D(
|
|
49
|
+
currentTensor,
|
|
50
|
+
weights.subsampleConv0Weight,
|
|
51
|
+
null,
|
|
52
|
+
{
|
|
53
|
+
inChannels: currentChannels,
|
|
54
|
+
outChannels: convChannels[0],
|
|
55
|
+
height: currentHeight,
|
|
56
|
+
width: currentWidth,
|
|
57
|
+
kernelH: 3,
|
|
58
|
+
kernelW: 3,
|
|
59
|
+
stride: 2,
|
|
60
|
+
pad: 1,
|
|
61
|
+
}
|
|
62
|
+
);
|
|
63
|
+
releaseBuffer(currentTensor.buffer);
|
|
64
|
+
currentHeight = Math.floor((currentHeight + 2 - 3) / 2) + 1;
|
|
65
|
+
currentWidth = Math.floor((currentWidth + 2 - 3) / 2) + 1;
|
|
66
|
+
currentChannels = convChannels[0];
|
|
67
|
+
|
|
68
|
+
// Norm after conv0
|
|
69
|
+
const norm0Elements = currentChannels * currentHeight * currentWidth;
|
|
70
|
+
const norm0Result = await runRMSNorm(
|
|
71
|
+
reshapeTensor(conv0Result, [norm0Elements, 1], 'audio_conv0_flat'),
|
|
72
|
+
weights.subsampleNorm0Weight,
|
|
73
|
+
audioConfig.rmsNormEps,
|
|
74
|
+
{ batchSize: currentHeight * currentWidth, hiddenSize: currentChannels }
|
|
75
|
+
);
|
|
76
|
+
releaseBuffer(conv0Result.buffer);
|
|
77
|
+
|
|
78
|
+
// Conv layer 1: [convChannels[0], H/2, W/2] -> [convChannels[1], H/4, W/4]
|
|
79
|
+
const conv1Input = reshapeTensor(norm0Result, [currentChannels, currentHeight, currentWidth], 'audio_conv1_in');
|
|
80
|
+
const conv1Result = await runConv2D(
|
|
81
|
+
conv1Input,
|
|
82
|
+
weights.subsampleConv1Weight,
|
|
83
|
+
null,
|
|
84
|
+
{
|
|
85
|
+
inChannels: currentChannels,
|
|
86
|
+
outChannels: convChannels[1],
|
|
87
|
+
height: currentHeight,
|
|
88
|
+
width: currentWidth,
|
|
89
|
+
kernelH: 3,
|
|
90
|
+
kernelW: 3,
|
|
91
|
+
stride: 2,
|
|
92
|
+
pad: 1,
|
|
93
|
+
}
|
|
94
|
+
);
|
|
95
|
+
releaseBuffer(norm0Result.buffer);
|
|
96
|
+
currentHeight = Math.floor((currentHeight + 2 - 3) / 2) + 1;
|
|
97
|
+
currentWidth = Math.floor((currentWidth + 2 - 3) / 2) + 1;
|
|
98
|
+
currentChannels = convChannels[1];
|
|
99
|
+
|
|
100
|
+
// Norm after conv1
|
|
101
|
+
const norm1Result = await runRMSNorm(
|
|
102
|
+
reshapeTensor(conv1Result, [currentHeight * currentWidth, currentChannels], 'audio_conv1_flat'),
|
|
103
|
+
weights.subsampleNorm1Weight,
|
|
104
|
+
audioConfig.rmsNormEps,
|
|
105
|
+
{ batchSize: currentHeight * currentWidth, hiddenSize: currentChannels }
|
|
106
|
+
);
|
|
107
|
+
releaseBuffer(conv1Result.buffer);
|
|
108
|
+
|
|
109
|
+
// Flatten to [seqLen, flatDim] where seqLen = currentHeight, flatDim = currentChannels * currentWidth
|
|
110
|
+
const seqLen = currentHeight;
|
|
111
|
+
const flatDim = currentChannels * currentWidth;
|
|
112
|
+
|
|
113
|
+
// Linear projection: [seqLen, flatDim] -> [seqLen, hiddenSize]
|
|
114
|
+
const projInput = reshapeTensor(norm1Result, [seqLen, flatDim], 'audio_subsample_flat');
|
|
115
|
+
const projected = await runMatmul(
|
|
116
|
+
projInput,
|
|
117
|
+
weights.subsampleInputProjWeight,
|
|
118
|
+
seqLen,
|
|
119
|
+
hiddenSize,
|
|
120
|
+
flatDim,
|
|
121
|
+
{ outputDtype: 'f32', transposeB: 'auto' }
|
|
122
|
+
);
|
|
123
|
+
releaseBuffer(norm1Result.buffer);
|
|
124
|
+
|
|
125
|
+
return { tensor: projected, seqLen };
|
|
126
|
+
}
|
|
127
|
+
|
|
128
|
+
// ---------------------------------------------------------------------------
|
|
129
|
+
// Conformer Half-Step FFN (Macaron style)
|
|
130
|
+
// ---------------------------------------------------------------------------
|
|
131
|
+
|
|
132
|
+
async function runConformerFFN(hiddenTensor, layerWeights, audioConfig, seqLen, prefix) {
|
|
133
|
+
const hiddenSize = audioConfig.hiddenSize;
|
|
134
|
+
const intermediateSize = hiddenSize * 4; // Conformer FFN is typically 4x
|
|
135
|
+
const residualWeight = audioConfig.residualWeight;
|
|
136
|
+
|
|
137
|
+
// Pre layer norm
|
|
138
|
+
const normed = await runRMSNorm(
|
|
139
|
+
hiddenTensor,
|
|
140
|
+
layerWeights.preLayerNorm,
|
|
141
|
+
audioConfig.rmsNormEps,
|
|
142
|
+
{ batchSize: seqLen, hiddenSize }
|
|
143
|
+
);
|
|
144
|
+
|
|
145
|
+
// FFW layer 1: [seqLen, hiddenSize] -> [seqLen, intermediateSize]
|
|
146
|
+
const ffn1 = await runClippableLinear(
|
|
147
|
+
normed,
|
|
148
|
+
layerWeights.ffwLayer1Weight,
|
|
149
|
+
seqLen,
|
|
150
|
+
intermediateSize,
|
|
151
|
+
hiddenSize,
|
|
152
|
+
layerWeights.ffwLayer1Clip,
|
|
153
|
+
`${prefix}_ffw1`
|
|
154
|
+
);
|
|
155
|
+
releaseBuffer(normed.buffer);
|
|
156
|
+
|
|
157
|
+
// SiLU activation
|
|
158
|
+
const activated = await runSiLU(ffn1, { size: seqLen * intermediateSize });
|
|
159
|
+
releaseBuffer(ffn1.buffer);
|
|
160
|
+
|
|
161
|
+
// FFW layer 2: [seqLen, intermediateSize] -> [seqLen, hiddenSize]
|
|
162
|
+
const ffn2 = await runClippableLinear(
|
|
163
|
+
activated,
|
|
164
|
+
layerWeights.ffwLayer2Weight,
|
|
165
|
+
seqLen,
|
|
166
|
+
hiddenSize,
|
|
167
|
+
intermediateSize,
|
|
168
|
+
layerWeights.ffwLayer2Clip,
|
|
169
|
+
`${prefix}_ffw2`
|
|
170
|
+
);
|
|
171
|
+
releaseBuffer(activated.buffer);
|
|
172
|
+
|
|
173
|
+
// Post layer norm
|
|
174
|
+
const postNormed = await runRMSNorm(
|
|
175
|
+
ffn2,
|
|
176
|
+
layerWeights.postLayerNorm,
|
|
177
|
+
audioConfig.rmsNormEps,
|
|
178
|
+
{ batchSize: seqLen, hiddenSize }
|
|
179
|
+
);
|
|
180
|
+
releaseBuffer(ffn2.buffer);
|
|
181
|
+
|
|
182
|
+
// Scale by residualWeight (0.5 for Macaron half-step)
|
|
183
|
+
if (residualWeight !== 1.0) {
|
|
184
|
+
const data = await readBuffer(postNormed.buffer, seqLen * hiddenSize * 4);
|
|
185
|
+
const f32 = new Float32Array(data instanceof ArrayBuffer ? data : data.buffer);
|
|
186
|
+
for (let i = 0; i < f32.length; i++) {
|
|
187
|
+
f32[i] *= residualWeight;
|
|
188
|
+
}
|
|
189
|
+
const device = getDevice();
|
|
190
|
+
device.queue.writeBuffer(postNormed.buffer, 0, f32);
|
|
191
|
+
}
|
|
192
|
+
|
|
193
|
+
// Residual add
|
|
194
|
+
const result = await runResidualAdd(hiddenTensor, postNormed, seqLen * hiddenSize);
|
|
195
|
+
releaseBuffer(postNormed.buffer);
|
|
196
|
+
|
|
197
|
+
return result;
|
|
198
|
+
}
|
|
199
|
+
|
|
200
|
+
// ---------------------------------------------------------------------------
|
|
201
|
+
// Conformer Self-Attention with Relative Position Encoding
|
|
202
|
+
// ---------------------------------------------------------------------------
|
|
203
|
+
|
|
204
|
+
async function runConformerAttention(hiddenTensor, layerWeights, audioConfig, seqLen) {
|
|
205
|
+
const hiddenSize = audioConfig.hiddenSize;
|
|
206
|
+
const numHeads = audioConfig.numAttentionHeads;
|
|
207
|
+
const headDim = audioConfig.headDim;
|
|
208
|
+
const logitCap = audioConfig.attentionLogitCap;
|
|
209
|
+
const contextLeft = audioConfig.attentionContextLeft;
|
|
210
|
+
const contextRight = audioConfig.attentionContextRight;
|
|
211
|
+
|
|
212
|
+
// Pre-attention norm
|
|
213
|
+
const normed = await runRMSNorm(
|
|
214
|
+
hiddenTensor,
|
|
215
|
+
layerWeights.normPreAttn,
|
|
216
|
+
audioConfig.rmsNormEps,
|
|
217
|
+
{ batchSize: seqLen, hiddenSize }
|
|
218
|
+
);
|
|
219
|
+
|
|
220
|
+
// Q/K/V projections with clipping
|
|
221
|
+
let qTensor = await runClippableLinear(
|
|
222
|
+
normed, layerWeights.qProj, seqLen, hiddenSize, hiddenSize,
|
|
223
|
+
layerWeights.qProjClip, 'audio_q_proj'
|
|
224
|
+
);
|
|
225
|
+
let kTensor = await runClippableLinear(
|
|
226
|
+
normed, layerWeights.kProj, seqLen, hiddenSize, hiddenSize,
|
|
227
|
+
layerWeights.kProjClip, 'audio_k_proj'
|
|
228
|
+
);
|
|
229
|
+
let vTensor = await runClippableLinear(
|
|
230
|
+
normed, layerWeights.vProj, seqLen, hiddenSize, hiddenSize,
|
|
231
|
+
layerWeights.vProjClip, 'audio_v_proj'
|
|
232
|
+
);
|
|
233
|
+
releaseBuffer(normed.buffer);
|
|
234
|
+
|
|
235
|
+
// Apply per_dim_scale to queries: Q *= softplus(per_dim_scale)
|
|
236
|
+
// per_dim_scale is [headDim], applied per-head
|
|
237
|
+
const perDimScaleData = await readBuffer(layerWeights.perDimScale, headDim * 4);
|
|
238
|
+
const perDimScaleF32 = new Float32Array(
|
|
239
|
+
perDimScaleData instanceof ArrayBuffer ? perDimScaleData : perDimScaleData.buffer
|
|
240
|
+
);
|
|
241
|
+
const qData = await readBuffer(qTensor.buffer, seqLen * hiddenSize * 4);
|
|
242
|
+
const qF32 = new Float32Array(qData instanceof ArrayBuffer ? qData : qData.buffer);
|
|
243
|
+
for (let s = 0; s < seqLen; s++) {
|
|
244
|
+
for (let h = 0; h < numHeads; h++) {
|
|
245
|
+
for (let d = 0; d < headDim; d++) {
|
|
246
|
+
const idx = s * hiddenSize + h * headDim + d;
|
|
247
|
+
const scale = Math.log(1 + Math.exp(perDimScaleF32[d])); // softplus
|
|
248
|
+
qF32[idx] *= scale;
|
|
249
|
+
}
|
|
250
|
+
}
|
|
251
|
+
}
|
|
252
|
+
|
|
253
|
+
// Compute attention scores: [seqLen, numHeads, seqLen]
|
|
254
|
+
// Score = Q @ K^T / sqrt(headDim)
|
|
255
|
+
const scaleFactor = 1.0 / Math.sqrt(headDim);
|
|
256
|
+
const scores = new Float32Array(numHeads * seqLen * seqLen);
|
|
257
|
+
|
|
258
|
+
const kData = await readBuffer(kTensor.buffer, seqLen * hiddenSize * 4);
|
|
259
|
+
const kF32 = new Float32Array(kData instanceof ArrayBuffer ? kData : kData.buffer);
|
|
260
|
+
|
|
261
|
+
for (let h = 0; h < numHeads; h++) {
|
|
262
|
+
for (let qi = 0; qi < seqLen; qi++) {
|
|
263
|
+
for (let ki = 0; ki < seqLen; ki++) {
|
|
264
|
+
let dot = 0;
|
|
265
|
+
for (let d = 0; d < headDim; d++) {
|
|
266
|
+
dot += qF32[qi * hiddenSize + h * headDim + d] * kF32[ki * hiddenSize + h * headDim + d];
|
|
267
|
+
}
|
|
268
|
+
scores[h * seqLen * seqLen + qi * seqLen + ki] = dot * scaleFactor;
|
|
269
|
+
}
|
|
270
|
+
}
|
|
271
|
+
}
|
|
272
|
+
|
|
273
|
+
// Add relative position bias via relative_k_proj
|
|
274
|
+
// This is a simplified approach: compute relative position embeddings and add to scores
|
|
275
|
+
const relKProjData = await readBuffer(layerWeights.relativeKProj, hiddenSize * hiddenSize * 4);
|
|
276
|
+
const relKProjF32 = new Float32Array(
|
|
277
|
+
relKProjData instanceof ArrayBuffer ? relKProjData : relKProjData.buffer
|
|
278
|
+
);
|
|
279
|
+
|
|
280
|
+
// Build relative position vectors and project
|
|
281
|
+
for (let h = 0; h < numHeads; h++) {
|
|
282
|
+
for (let qi = 0; qi < seqLen; qi++) {
|
|
283
|
+
for (let ki = 0; ki < seqLen; ki++) {
|
|
284
|
+
const relPos = ki - qi;
|
|
285
|
+
// Compute relative bias from q and relative_k_proj
|
|
286
|
+
let relBias = 0;
|
|
287
|
+
for (let d = 0; d < headDim; d++) {
|
|
288
|
+
const qVal = qF32[qi * hiddenSize + h * headDim + d];
|
|
289
|
+
// Use distance-based relative encoding
|
|
290
|
+
relBias += qVal * relKProjF32[(h * headDim + d) * hiddenSize + (Math.abs(relPos) % hiddenSize)];
|
|
291
|
+
}
|
|
292
|
+
scores[h * seqLen * seqLen + qi * seqLen + ki] += relBias * scaleFactor;
|
|
293
|
+
}
|
|
294
|
+
}
|
|
295
|
+
}
|
|
296
|
+
|
|
297
|
+
// Apply logit capping: scores = cap * tanh(scores / cap)
|
|
298
|
+
for (let i = 0; i < scores.length; i++) {
|
|
299
|
+
scores[i] = logitCap * Math.tanh(scores[i] / logitCap);
|
|
300
|
+
}
|
|
301
|
+
|
|
302
|
+
// Apply chunked attention mask (contextLeft, contextRight)
|
|
303
|
+
const invalidValue = audioConfig.attentionInvalidLogitsValue;
|
|
304
|
+
for (let h = 0; h < numHeads; h++) {
|
|
305
|
+
for (let qi = 0; qi < seqLen; qi++) {
|
|
306
|
+
for (let ki = 0; ki < seqLen; ki++) {
|
|
307
|
+
const dist = ki - qi;
|
|
308
|
+
if (dist < -contextLeft || dist > contextRight) {
|
|
309
|
+
scores[h * seqLen * seqLen + qi * seqLen + ki] = invalidValue;
|
|
310
|
+
}
|
|
311
|
+
}
|
|
312
|
+
}
|
|
313
|
+
}
|
|
314
|
+
|
|
315
|
+
// Softmax per query position per head
|
|
316
|
+
for (let h = 0; h < numHeads; h++) {
|
|
317
|
+
for (let qi = 0; qi < seqLen; qi++) {
|
|
318
|
+
const offset = h * seqLen * seqLen + qi * seqLen;
|
|
319
|
+
let maxVal = -Infinity;
|
|
320
|
+
for (let ki = 0; ki < seqLen; ki++) {
|
|
321
|
+
if (scores[offset + ki] > maxVal) maxVal = scores[offset + ki];
|
|
322
|
+
}
|
|
323
|
+
let sumExp = 0;
|
|
324
|
+
for (let ki = 0; ki < seqLen; ki++) {
|
|
325
|
+
scores[offset + ki] = Math.exp(scores[offset + ki] - maxVal);
|
|
326
|
+
sumExp += scores[offset + ki];
|
|
327
|
+
}
|
|
328
|
+
for (let ki = 0; ki < seqLen; ki++) {
|
|
329
|
+
scores[offset + ki] /= sumExp;
|
|
330
|
+
}
|
|
331
|
+
}
|
|
332
|
+
}
|
|
333
|
+
|
|
334
|
+
// Apply attention: output = scores @ V
|
|
335
|
+
const vData = await readBuffer(vTensor.buffer, seqLen * hiddenSize * 4);
|
|
336
|
+
const vF32 = new Float32Array(vData instanceof ArrayBuffer ? vData : vData.buffer);
|
|
337
|
+
const attnOutput = new Float32Array(seqLen * hiddenSize);
|
|
338
|
+
|
|
339
|
+
for (let h = 0; h < numHeads; h++) {
|
|
340
|
+
for (let qi = 0; qi < seqLen; qi++) {
|
|
341
|
+
for (let d = 0; d < headDim; d++) {
|
|
342
|
+
let sum = 0;
|
|
343
|
+
for (let ki = 0; ki < seqLen; ki++) {
|
|
344
|
+
sum += scores[h * seqLen * seqLen + qi * seqLen + ki] * vF32[ki * hiddenSize + h * headDim + d];
|
|
345
|
+
}
|
|
346
|
+
attnOutput[qi * hiddenSize + h * headDim + d] = sum;
|
|
347
|
+
}
|
|
348
|
+
}
|
|
349
|
+
}
|
|
350
|
+
|
|
351
|
+
releaseBuffer(qTensor.buffer);
|
|
352
|
+
releaseBuffer(kTensor.buffer);
|
|
353
|
+
releaseBuffer(vTensor.buffer);
|
|
354
|
+
|
|
355
|
+
// Upload attention output
|
|
356
|
+
const attnBuffer = await uploadData(attnOutput, 'audio_attn_output');
|
|
357
|
+
const attnTensor = createTensor(attnBuffer, 'f32', [seqLen, hiddenSize], 'audio_attn_output');
|
|
358
|
+
|
|
359
|
+
// Output projection with clipping
|
|
360
|
+
const projected = await runClippableLinear(
|
|
361
|
+
attnTensor,
|
|
362
|
+
layerWeights.postProj,
|
|
363
|
+
seqLen,
|
|
364
|
+
hiddenSize,
|
|
365
|
+
hiddenSize,
|
|
366
|
+
layerWeights.postProjClip,
|
|
367
|
+
'audio_attn_post'
|
|
368
|
+
);
|
|
369
|
+
releaseBuffer(attnTensor.buffer);
|
|
370
|
+
|
|
371
|
+
// Post-attention norm
|
|
372
|
+
const postNormed = await runRMSNorm(
|
|
373
|
+
projected,
|
|
374
|
+
layerWeights.normPostAttn,
|
|
375
|
+
audioConfig.rmsNormEps,
|
|
376
|
+
{ batchSize: seqLen, hiddenSize }
|
|
377
|
+
);
|
|
378
|
+
releaseBuffer(projected.buffer);
|
|
379
|
+
|
|
380
|
+
// Residual add
|
|
381
|
+
const result = await runResidualAdd(hiddenTensor, postNormed, seqLen * hiddenSize);
|
|
382
|
+
releaseBuffer(postNormed.buffer);
|
|
383
|
+
|
|
384
|
+
return result;
|
|
385
|
+
}
|
|
386
|
+
|
|
387
|
+
// ---------------------------------------------------------------------------
|
|
388
|
+
// Conformer Convolution Module (LConv1D)
|
|
389
|
+
// ---------------------------------------------------------------------------
|
|
390
|
+
|
|
391
|
+
async function runConformerConvModule(hiddenTensor, layerWeights, audioConfig, seqLen) {
|
|
392
|
+
const hiddenSize = audioConfig.hiddenSize;
|
|
393
|
+
const kernelSize = audioConfig.convKernelSize;
|
|
394
|
+
|
|
395
|
+
// Pre layer norm
|
|
396
|
+
const normed = await runRMSNorm(
|
|
397
|
+
hiddenTensor,
|
|
398
|
+
layerWeights.lconvPreLayerNorm,
|
|
399
|
+
audioConfig.rmsNormEps,
|
|
400
|
+
{ batchSize: seqLen, hiddenSize }
|
|
401
|
+
);
|
|
402
|
+
|
|
403
|
+
// Linear start (pointwise): [seqLen, hiddenSize] -> [seqLen, hiddenSize*2]
|
|
404
|
+
// The GLU split: first half goes through sigmoid gate, second half is the value
|
|
405
|
+
const linearStart = await runClippableLinear(
|
|
406
|
+
normed,
|
|
407
|
+
layerWeights.lconvLinearStartWeight,
|
|
408
|
+
seqLen,
|
|
409
|
+
hiddenSize * 2,
|
|
410
|
+
hiddenSize,
|
|
411
|
+
layerWeights.lconvLinearStartClip,
|
|
412
|
+
'audio_lconv_start'
|
|
413
|
+
);
|
|
414
|
+
releaseBuffer(normed.buffer);
|
|
415
|
+
|
|
416
|
+
// GLU: split into gate and value, gate = sigmoid(gate_half), output = gate * value_half
|
|
417
|
+
const startData = await readBuffer(linearStart.buffer, seqLen * hiddenSize * 2 * 4);
|
|
418
|
+
const startF32 = new Float32Array(startData instanceof ArrayBuffer ? startData : startData.buffer);
|
|
419
|
+
const gluOutput = new Float32Array(seqLen * hiddenSize);
|
|
420
|
+
for (let s = 0; s < seqLen; s++) {
|
|
421
|
+
for (let d = 0; d < hiddenSize; d++) {
|
|
422
|
+
const gateVal = startF32[s * hiddenSize * 2 + d];
|
|
423
|
+
const valueVal = startF32[s * hiddenSize * 2 + hiddenSize + d];
|
|
424
|
+
const sigmoid = 1.0 / (1.0 + Math.exp(-gateVal));
|
|
425
|
+
gluOutput[s * hiddenSize + d] = sigmoid * valueVal;
|
|
426
|
+
}
|
|
427
|
+
}
|
|
428
|
+
releaseBuffer(linearStart.buffer);
|
|
429
|
+
|
|
430
|
+
// Depthwise Conv1D: [channels=hiddenSize, length=seqLen]
|
|
431
|
+
// Transpose from [seqLen, hiddenSize] to [hiddenSize, seqLen]
|
|
432
|
+
const transposed = new Float32Array(hiddenSize * seqLen);
|
|
433
|
+
for (let s = 0; s < seqLen; s++) {
|
|
434
|
+
for (let d = 0; d < hiddenSize; d++) {
|
|
435
|
+
transposed[d * seqLen + s] = gluOutput[s * hiddenSize + d];
|
|
436
|
+
}
|
|
437
|
+
}
|
|
438
|
+
const transposedBuffer = await uploadData(transposed, 'audio_lconv_transposed');
|
|
439
|
+
const transposedTensor = createTensor(transposedBuffer, 'f32', [hiddenSize, seqLen], 'audio_lconv_transposed');
|
|
440
|
+
|
|
441
|
+
const convResult = await runDepthwiseConv1D(
|
|
442
|
+
transposedTensor,
|
|
443
|
+
layerWeights.lconvDepthwiseWeight,
|
|
444
|
+
{ channels: hiddenSize, length: seqLen, kernelSize }
|
|
445
|
+
);
|
|
446
|
+
releaseBuffer(transposedTensor.buffer);
|
|
447
|
+
|
|
448
|
+
// Transpose back: [hiddenSize, seqLen] -> [seqLen, hiddenSize]
|
|
449
|
+
const convData = await readBuffer(convResult.buffer, hiddenSize * seqLen * 4);
|
|
450
|
+
const convF32 = new Float32Array(convData instanceof ArrayBuffer ? convData : convData.buffer);
|
|
451
|
+
const convTransposed = new Float32Array(seqLen * hiddenSize);
|
|
452
|
+
for (let s = 0; s < seqLen; s++) {
|
|
453
|
+
for (let d = 0; d < hiddenSize; d++) {
|
|
454
|
+
convTransposed[s * hiddenSize + d] = convF32[d * seqLen + s];
|
|
455
|
+
}
|
|
456
|
+
}
|
|
457
|
+
releaseBuffer(convResult.buffer);
|
|
458
|
+
|
|
459
|
+
// Conv norm (RMSNorm on the convolution output)
|
|
460
|
+
const convNormBuffer = await uploadData(convTransposed, 'audio_lconv_normed');
|
|
461
|
+
const convNormTensor = createTensor(convNormBuffer, 'f32', [seqLen, hiddenSize], 'audio_lconv_normed');
|
|
462
|
+
const convNormed = await runRMSNorm(
|
|
463
|
+
convNormTensor,
|
|
464
|
+
layerWeights.lconvConvNorm,
|
|
465
|
+
audioConfig.rmsNormEps,
|
|
466
|
+
{ batchSize: seqLen, hiddenSize }
|
|
467
|
+
);
|
|
468
|
+
releaseBuffer(convNormTensor.buffer);
|
|
469
|
+
|
|
470
|
+
// SiLU activation
|
|
471
|
+
const activated = await runSiLU(convNormed, { size: seqLen * hiddenSize });
|
|
472
|
+
releaseBuffer(convNormed.buffer);
|
|
473
|
+
|
|
474
|
+
// Linear end (pointwise): [seqLen, hiddenSize] -> [seqLen, hiddenSize]
|
|
475
|
+
const linearEnd = await runClippableLinear(
|
|
476
|
+
activated,
|
|
477
|
+
layerWeights.lconvLinearEndWeight,
|
|
478
|
+
seqLen,
|
|
479
|
+
hiddenSize,
|
|
480
|
+
hiddenSize,
|
|
481
|
+
layerWeights.lconvLinearEndClip,
|
|
482
|
+
'audio_lconv_end'
|
|
483
|
+
);
|
|
484
|
+
releaseBuffer(activated.buffer);
|
|
485
|
+
|
|
486
|
+
// Residual add
|
|
487
|
+
const result = await runResidualAdd(hiddenTensor, linearEnd, seqLen * hiddenSize);
|
|
488
|
+
releaseBuffer(linearEnd.buffer);
|
|
489
|
+
|
|
490
|
+
return result;
|
|
491
|
+
}
|
|
492
|
+
|
|
493
|
+
// ---------------------------------------------------------------------------
|
|
494
|
+
// Full Conformer Layer
|
|
495
|
+
// ---------------------------------------------------------------------------
|
|
496
|
+
|
|
497
|
+
async function runConformerLayer(hiddenTensor, layerWeights, audioConfig, seqLen, layerIdx) {
|
|
498
|
+
// 1. Half-step FFN1 (Macaron)
|
|
499
|
+
let current = await runConformerFFN(
|
|
500
|
+
hiddenTensor,
|
|
501
|
+
layerWeights.feedForward1,
|
|
502
|
+
audioConfig,
|
|
503
|
+
seqLen,
|
|
504
|
+
`audio_layer${layerIdx}_ff1`
|
|
505
|
+
);
|
|
506
|
+
|
|
507
|
+
// 2. Self-attention with relative position encoding
|
|
508
|
+
const afterAttn = await runConformerAttention(
|
|
509
|
+
current,
|
|
510
|
+
layerWeights,
|
|
511
|
+
audioConfig,
|
|
512
|
+
seqLen
|
|
513
|
+
);
|
|
514
|
+
releaseBuffer(current.buffer);
|
|
515
|
+
current = afterAttn;
|
|
516
|
+
|
|
517
|
+
// 3. Convolution module
|
|
518
|
+
const afterConv = await runConformerConvModule(
|
|
519
|
+
current,
|
|
520
|
+
layerWeights,
|
|
521
|
+
audioConfig,
|
|
522
|
+
seqLen
|
|
523
|
+
);
|
|
524
|
+
releaseBuffer(current.buffer);
|
|
525
|
+
current = afterConv;
|
|
526
|
+
|
|
527
|
+
// 4. Half-step FFN2
|
|
528
|
+
const afterFFN2 = await runConformerFFN(
|
|
529
|
+
current,
|
|
530
|
+
layerWeights.feedForward2,
|
|
531
|
+
audioConfig,
|
|
532
|
+
seqLen,
|
|
533
|
+
`audio_layer${layerIdx}_ff2`
|
|
534
|
+
);
|
|
535
|
+
releaseBuffer(current.buffer);
|
|
536
|
+
current = afterFFN2;
|
|
537
|
+
|
|
538
|
+
// 5. Final norm_out
|
|
539
|
+
const hiddenSize = audioConfig.hiddenSize;
|
|
540
|
+
const normed = await runRMSNorm(
|
|
541
|
+
current,
|
|
542
|
+
layerWeights.normOut,
|
|
543
|
+
audioConfig.rmsNormEps,
|
|
544
|
+
{ batchSize: seqLen, hiddenSize }
|
|
545
|
+
);
|
|
546
|
+
releaseBuffer(current.buffer);
|
|
547
|
+
|
|
548
|
+
return normed;
|
|
549
|
+
}
|
|
550
|
+
|
|
551
|
+
// ---------------------------------------------------------------------------
|
|
552
|
+
// Main Encoder Entry Point
|
|
553
|
+
// ---------------------------------------------------------------------------
|
|
554
|
+
|
|
555
|
+
/**
|
|
556
|
+
* Encode audio through the Gemma 4 conformer audio tower.
|
|
557
|
+
*
|
|
558
|
+
* @param {object} params
|
|
559
|
+
* @param {Float32Array} params.melFeatures Log-mel spectrogram [numFrames * nMels]
|
|
560
|
+
* @param {number} params.numFrames Number of mel frames
|
|
561
|
+
* @param {number} params.nMels Number of mel bands
|
|
562
|
+
* @param {object} params.audioConfig Resolved audio encoder config
|
|
563
|
+
* @param {object} params.weights Audio encoder weight buffers
|
|
564
|
+
* @returns {Promise<{ features: GPUBuffer, numTokens: number }>}
|
|
565
|
+
*/
|
|
566
|
+
export async function encodeGemma4Audio(params) {
|
|
567
|
+
const { melFeatures, numFrames, nMels, audioConfig, weights } = params;
|
|
568
|
+
|
|
569
|
+
if (audioConfig.audioArchitecture !== 'gemma4') {
|
|
570
|
+
throw new Error(
|
|
571
|
+
`[Audio] Unsupported audio architecture "${audioConfig.audioArchitecture}". Expected "gemma4".`
|
|
572
|
+
);
|
|
573
|
+
}
|
|
574
|
+
if (!audioConfig.useClippedLinears) {
|
|
575
|
+
throw new Error('[Audio] Gemma 4 audio encoder requires useClippedLinears=true.');
|
|
576
|
+
}
|
|
577
|
+
|
|
578
|
+
const hiddenSize = audioConfig.hiddenSize;
|
|
579
|
+
const depth = audioConfig.depth;
|
|
580
|
+
|
|
581
|
+
log.debug('Audio', `encodeGemma4Audio: ${numFrames} frames, ${nMels} mels, ${depth} layers`);
|
|
582
|
+
|
|
583
|
+
// Step 1: Subsampling — Conv -> Norm -> Linear projection
|
|
584
|
+
const { tensor: subsampledTensor, seqLen } = await runSubsampling(
|
|
585
|
+
melFeatures, numFrames, nMels, weights, audioConfig
|
|
586
|
+
);
|
|
587
|
+
|
|
588
|
+
log.debug('Audio', `Subsampling complete: seqLen=${seqLen}, hiddenSize=${hiddenSize}`);
|
|
589
|
+
|
|
590
|
+
// Step 2: Conformer layers
|
|
591
|
+
let currentTensor = subsampledTensor;
|
|
592
|
+
for (let i = 0; i < depth; i++) {
|
|
593
|
+
const layerResult = await runConformerLayer(
|
|
594
|
+
currentTensor,
|
|
595
|
+
weights.layers[i],
|
|
596
|
+
audioConfig,
|
|
597
|
+
seqLen,
|
|
598
|
+
i
|
|
599
|
+
);
|
|
600
|
+
if (currentTensor !== subsampledTensor || i > 0) {
|
|
601
|
+
releaseBuffer(currentTensor.buffer);
|
|
602
|
+
}
|
|
603
|
+
currentTensor = layerResult;
|
|
604
|
+
}
|
|
605
|
+
if (depth > 0) {
|
|
606
|
+
// subsampledTensor was released in the loop
|
|
607
|
+
} else {
|
|
608
|
+
releaseBuffer(subsampledTensor.buffer);
|
|
609
|
+
}
|
|
610
|
+
|
|
611
|
+
// Step 3: Output projection [seqLen, hiddenSize] -> [seqLen, outputProjDims]
|
|
612
|
+
const outputProjDims = audioConfig.outputProjDims;
|
|
613
|
+
const outputProj = await runMatmul(
|
|
614
|
+
currentTensor,
|
|
615
|
+
weights.outputProjWeight,
|
|
616
|
+
seqLen,
|
|
617
|
+
outputProjDims,
|
|
618
|
+
hiddenSize,
|
|
619
|
+
{ outputDtype: 'f32', transposeB: 'auto' }
|
|
620
|
+
);
|
|
621
|
+
releaseBuffer(currentTensor.buffer);
|
|
622
|
+
|
|
623
|
+
// Add output projection bias
|
|
624
|
+
if (weights.outputProjBias) {
|
|
625
|
+
const biasData = await readBuffer(weights.outputProjBias, outputProjDims * 4);
|
|
626
|
+
const biasF32 = new Float32Array(biasData instanceof ArrayBuffer ? biasData : biasData.buffer);
|
|
627
|
+
const projData = await readBuffer(outputProj.buffer, seqLen * outputProjDims * 4);
|
|
628
|
+
const projF32 = new Float32Array(projData instanceof ArrayBuffer ? projData : projData.buffer);
|
|
629
|
+
for (let s = 0; s < seqLen; s++) {
|
|
630
|
+
for (let d = 0; d < outputProjDims; d++) {
|
|
631
|
+
projF32[s * outputProjDims + d] += biasF32[d];
|
|
632
|
+
}
|
|
633
|
+
}
|
|
634
|
+
const device = getDevice();
|
|
635
|
+
device.queue.writeBuffer(outputProj.buffer, 0, projF32);
|
|
636
|
+
}
|
|
637
|
+
|
|
638
|
+
// Step 4: Audio embedding projection [seqLen, outputProjDims] -> [seqLen, outputProjDims]
|
|
639
|
+
const embedProj = await runMatmul(
|
|
640
|
+
outputProj,
|
|
641
|
+
weights.audioEmbeddingProjWeight,
|
|
642
|
+
seqLen,
|
|
643
|
+
outputProjDims,
|
|
644
|
+
outputProjDims,
|
|
645
|
+
{ outputDtype: 'f32', transposeB: 'auto' }
|
|
646
|
+
);
|
|
647
|
+
releaseBuffer(outputProj.buffer);
|
|
648
|
+
|
|
649
|
+
log.info('Audio', `Gemma 4 audio encoding complete: ${seqLen} tokens, ${outputProjDims} dims`);
|
|
650
|
+
|
|
651
|
+
return {
|
|
652
|
+
features: embedProj.buffer,
|
|
653
|
+
numTokens: seqLen,
|
|
654
|
+
};
|
|
655
|
+
}
|
|
@@ -0,0 +1,39 @@
|
|
|
1
|
+
|
|
2
|
+
import { log } from '../../../debug/index.js';
|
|
3
|
+
import { encodeGemma4Audio } from './gemma4.js';
|
|
4
|
+
|
|
5
|
+
/**
|
|
6
|
+
* Encode audio through the audio pipeline.
|
|
7
|
+
*
|
|
8
|
+
* Routes to architecture-specific encoder based on audioConfig.audioArchitecture.
|
|
9
|
+
*
|
|
10
|
+
* @param {object} params
|
|
11
|
+
* @param {Float32Array} params.melFeatures Log-mel spectrogram [numFrames * nMels]
|
|
12
|
+
* @param {number} params.numFrames Number of mel frames
|
|
13
|
+
* @param {number} params.nMels Number of mel bands
|
|
14
|
+
* @param {object} params.audioConfig Audio encoder config from manifest
|
|
15
|
+
* @param {object} params.weights Audio encoder weight buffers
|
|
16
|
+
* @returns {Promise<AudioEncodeResult>}
|
|
17
|
+
*/
|
|
18
|
+
export async function encodeAudio(params) {
|
|
19
|
+
const { audioConfig } = params;
|
|
20
|
+
const arch = audioConfig?.audioArchitecture ?? null;
|
|
21
|
+
|
|
22
|
+
log.debug('Audio', `encodeAudio: arch=${arch}`);
|
|
23
|
+
|
|
24
|
+
switch (arch) {
|
|
25
|
+
case 'gemma4':
|
|
26
|
+
return encodeGemma4Audio(params);
|
|
27
|
+
default:
|
|
28
|
+
throw new Error(
|
|
29
|
+
`Unsupported audio architecture "${arch}". ` +
|
|
30
|
+
'Supported: gemma4. Check audio_config.audio_architecture.'
|
|
31
|
+
);
|
|
32
|
+
}
|
|
33
|
+
}
|
|
34
|
+
|
|
35
|
+
/**
|
|
36
|
+
* @typedef {object} AudioEncodeResult
|
|
37
|
+
* @property {GPUBuffer} features Encoded audio tokens [numTokens, outputDims]
|
|
38
|
+
* @property {number} numTokens Number of audio tokens after subsampling
|
|
39
|
+
*/
|