doppler-gpu 0.4.2 → 0.4.4
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 +210 -0
- package/README.md +63 -38
- package/assets/doppler.svg +57 -1601
- package/package.json +111 -32
- package/src/cli/cli-model-resolution.js +275 -0
- package/src/cli/cli-output.js +486 -0
- package/src/cli/doppler-cli.d.ts +18 -0
- package/src/cli/doppler-cli.js +885 -744
- package/src/cli/doppler-quickstart.d.ts +7 -23
- package/src/cli/doppler-quickstart.js +4 -2
- package/src/cli/doppler-serve.d.ts +210 -0
- package/src/cli/doppler-serve.js +469 -43
- package/src/cli/runtime-profiles.d.ts +40 -0
- package/src/cli/runtime-profiles.js +225 -0
- package/src/client/doppler-api.js +1 -6
- package/src/client/doppler-registry.d.ts +12 -0
- package/src/client/doppler-registry.js +27 -0
- package/src/client/doppler-registry.json +66 -10
- package/src/client/failure-taxonomy.js +139 -0
- package/src/client/fault-injection.js +30 -0
- package/src/client/provider.d.ts +151 -0
- package/src/client/provider.js +374 -0
- package/src/client/receipt.js +156 -0
- package/src/client/runtime/index.js +69 -4
- package/src/client/runtime/lora.d.ts +11 -1
- package/src/client/runtime/lora.js +28 -7
- package/src/client/runtime/model-manager.d.ts +0 -14
- package/src/client/runtime/model-manager.js +34 -50
- package/src/client/runtime/model-session.d.ts +2 -1
- package/src/client/runtime/model-session.js +35 -2
- package/src/client/runtime/model-source.d.ts +13 -1
- package/src/client/runtime/model-source.js +148 -3
- package/src/client/runtime/node-quickstart-cache.d.ts +22 -0
- package/src/client/runtime/node-quickstart-cache.js +353 -0
- package/src/client/runtime/source-runtime.js +1 -10
- package/src/client/runtime/types.d.ts +8 -60
- package/src/client/wrap-pipeline-handle.d.ts +82 -0
- package/src/client/wrap-pipeline-handle.js +259 -0
- package/src/config/README.md +3 -1
- package/src/config/conversion/README.md +28 -157
- package/src/config/conversion/diffusiongemma/diffusiongemma-26b-a4b-it-q4k-ehf16-af16.json +646 -0
- package/src/config/conversion/embeddinggemma/google-embeddinggemma-300m-q4k-ehf16-af32.json +21 -6
- package/src/config/conversion/gemma3/gemma-3-1b-it-f16-af32.json +19 -12
- package/src/config/conversion/gemma3/gemma-3-1b-it-q4k-ehf16-af32.json +27 -12
- package/src/config/conversion/gemma3/gemma-3-270m-it-f16-af32.json +393 -0
- package/src/config/conversion/gemma3/gemma-3-270m-it-q4k-ehf16-af16.json +425 -0
- package/src/config/conversion/gemma3/gemma-3-270m-it-q4k-ehf16-af32.json +39 -16
- package/src/config/conversion/gemma3/translategemma-4b-1b-enes-q4k-ehf16-af32.json +19 -12
- package/src/config/conversion/gemma3/translategemma-4b-it-q4k-ehf16-af32.json +27 -12
- package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-af16.json +737 -0
- package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-af32.json +730 -0
- package/src/config/conversion/gemma4/gemma-4-12b-it-text-q4k-ehf16-hq4k-af16.json +780 -0
- package/src/config/conversion/gemma4/gemma-4-12b-it-text-w4a16-ct-ehf16-af16.json +716 -0
- package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af16.json +719 -0
- package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af32.json +715 -0
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af16-int4ple.json +591 -0
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32-int4ple.json +566 -0
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32.json +367 -61
- package/src/config/conversion/gemma4/gemma-4-moe-q4k-ehf16-af32.json +212 -55
- package/src/config/conversion/gpt-oss-20b-f16-xmxfp4.json +216 -57
- package/src/config/conversion/janus/janus-pro-1b-text-q4k-ehaf16.json +229 -57
- package/src/config/conversion/lfm2/lfm2.5-1.2b-instruct-q4k-ehf16-af32.json +19 -12
- package/src/config/conversion/qwen3/qwen-3-5-0-8b-q4k-ehaf16.json +82 -36
- package/src/config/conversion/qwen3/qwen-3-5-2b-q4k-ehaf16.json +96 -60
- package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-eaf16.json +521 -0
- package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-ehaf16.json +515 -0
- package/src/config/execution-contract-check.js +6 -5
- package/src/config/goal-completion-matrix.json +552 -0
- package/src/config/kernel-path-loader.d.ts +8 -2
- package/src/config/kernel-path-loader.js +192 -38
- package/src/config/kernels/kernel-ref-digests.js +163 -126
- package/src/config/kernels/registry.d.ts +0 -35
- package/src/config/kernels/registry.js +5 -9
- package/src/config/kernels/registry.json +3593 -1146
- package/src/config/merge-contract-check.js +7 -3
- package/src/config/merge-helpers.d.ts +3 -1
- package/src/config/merge-helpers.js +2 -1
- package/src/config/merge.d.ts +1 -0
- package/src/config/merge.js +65 -7
- package/src/config/model-lanes/catalog-lane-resolver.d.ts +66 -0
- package/src/config/model-lanes/catalog-lane-resolver.js +170 -0
- package/src/config/param-categories.d.ts +0 -9
- package/src/config/param-categories.js +1 -5
- package/src/config/param-validator.d.ts +2 -0
- package/src/config/param-validator.js +20 -0
- package/src/config/platforms/loader.d.ts +0 -27
- package/src/config/platforms/loader.js +0 -30
- package/src/config/required-inference-fields-contract-check.js +1 -7
- package/src/config/runtime/README.md +28 -0
- package/src/config/runtime/experiments/bench/gemma4-bench-q4k.json +1 -1
- package/src/config/runtime/experiments/bench/qwen3-bench-q4k-batched.json +2 -2
- package/src/config/runtime/experiments/debug/gemma4-e2b-browser-diagnose-q4k.json +85 -0
- package/src/config/runtime/experiments/verify/lfm2-verify.json +3 -3
- package/src/config/runtime/platform/metal-apple-q4k.json +19 -3
- package/src/config/runtime/profiles/diffusiongemma-26b-a4b-throughput.json +26 -0
- package/src/config/runtime/profiles/gemma3-270m-q4k-throughput-overlapped-probe.json +42 -0
- package/src/config/runtime/profiles/gemma3-270m-q4k-throughput-probe.json +39 -0
- package/src/config/runtime/profiles/gemma4-12b-rdrr-safe-b4.json +48 -0
- package/src/config/runtime/profiles/gemma4-12b-rdrr-safe-single.json +46 -0
- package/src/config/runtime/profiles/gemma4-31b-f16-activations-probe.json +65 -0
- package/src/config/runtime/profiles/gemma4-e2b-baseline-probe.json +41 -0
- package/src/config/runtime/profiles/gemma4-e2b-fused-rmsnorm-probe.json +44 -0
- package/src/config/runtime/profiles/gemma4-e2b-fused-rmsnorm-test.json +19 -0
- package/src/config/runtime/profiles/gemma4-e2b-int4ple-baseline-probe.json +31 -0
- package/src/config/runtime/profiles/gemma4-e2b-int4ple-f16a-probe.json +34 -0
- package/src/config/runtime/profiles/gemma4-e2b-int4ple-reference-export.json +64 -0
- package/src/config/runtime/profiles/gemma4-e2b-int4ple-retain-q4k-probe.json +47 -0
- package/src/config/runtime/profiles/gemma4-e2b-prefill-profile.json +35 -0
- package/src/config/runtime/profiles/gemma4-e2b-throughput-probe.json +16 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-prefill-profile.json +35 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-throughput-probe.json +16 -0
- package/src/config/runtime/profiles/qwen3-6-27b-f16-activations-probe.json +62 -0
- package/src/config/schema/bridge.schema.d.ts +16 -0
- package/src/config/schema/browser-suite-metrics.schema.d.ts +1 -0
- package/src/config/schema/browser-suite-metrics.schema.js +2 -0
- package/src/config/schema/buffer-pool.schema.js +3 -3
- package/src/config/schema/claim-evidence-contract.schema.json +93 -0
- package/src/config/schema/command-surface-contract.schema.json +81 -0
- package/src/config/schema/converter.schema.d.ts +16 -0
- package/src/config/schema/converter.schema.js +11 -0
- package/src/config/schema/debug.schema.d.ts +22 -1
- package/src/config/schema/debug.schema.js +1 -0
- package/src/config/schema/distill-training.schema.d.ts +0 -1
- package/src/config/schema/distill-training.schema.js +1 -1
- package/src/config/schema/doppler.schema.d.ts +13 -2
- package/src/config/schema/doppler.schema.js +140 -8
- package/src/config/schema/emulation.schema.d.ts +3 -0
- package/src/config/schema/emulation.schema.js +16 -16
- package/src/config/schema/execution-v1.schema.d.ts +39 -1
- package/src/config/schema/execution-v1.schema.js +35 -20
- package/src/config/schema/goal-completion-matrix.schema.json +223 -0
- package/src/config/schema/index.d.ts +52 -1
- package/src/config/schema/index.js +14 -0
- package/src/config/schema/inference.schema.d.ts +14 -3
- package/src/config/schema/kernel-path.schema.d.ts +6 -27
- package/src/config/schema/kernel-thresholds.schema.d.ts +0 -2
- package/src/config/schema/kernel-thresholds.schema.js +3 -7
- package/src/config/schema/kvcache.schema.js +0 -9
- package/src/config/schema/loading.schema.d.ts +12 -0
- package/src/config/schema/loading.schema.js +4 -0
- package/src/config/schema/manifest.schema.d.ts +175 -2
- package/src/config/schema/manifest.schema.js +58 -3
- package/src/config/schema/model-artifact-contract.schema.json +119 -0
- package/src/config/schema/moe.schema.d.ts +13 -0
- package/src/config/schema/moe.schema.js +1 -9
- package/src/config/schema/policy-schema-registry.json +26 -0
- package/src/config/schema/program-bundle.schema.d.ts +233 -0
- package/src/config/schema/program-bundle.schema.js +393 -0
- package/src/config/schema/subsystem-support-contract.schema.json +48 -0
- package/src/config/schema/training-metrics.schema.d.ts +1 -1
- package/src/config/schema/training-metrics.schema.js +4 -3
- package/src/config/schema/ul-training.schema.d.ts +0 -1
- package/src/config/schema/ul-training.schema.js +1 -1
- package/src/config/source-packages/litert/gemma-4-12b-it.json +390 -0
- package/src/config/source-packages/litert/gemma-4-e2b-it.json +41 -20
- package/src/config/source-packages/litert/gemma-4-e4b-it.json +370 -0
- package/src/config/source-packages/registry.json +25 -0
- package/src/config/supported-operations.d.ts +14 -0
- package/src/config/supported-operations.js +34 -0
- package/src/config/supported-operations.json +183 -0
- package/src/config/training-defaults.js +1 -7
- package/src/config/transforms/capability-transform-resolver.d.ts +13 -2
- package/src/config/transforms/capability-transform-resolver.js +3 -1
- package/src/config/transforms/execution-graph-transforms.d.ts +97 -0
- package/src/config/transforms/execution-graph-transforms.js +829 -168
- package/src/converter/conversion-plan.d.ts +2 -0
- package/src/converter/conversion-plan.js +31 -14
- package/src/converter/core.d.ts +65 -0
- package/src/converter/core.js +1153 -22
- package/src/converter/parsers/diffusion.js +1 -33
- package/src/converter/quantization-info.d.ts +5 -4
- package/src/converter/quantization-info.js +218 -5
- package/src/converter/quantizer.d.ts +24 -4
- package/src/converter/quantizer.js +92 -2
- package/src/debug/config.d.ts +0 -4
- package/src/debug/config.js +0 -17
- package/src/debug/index.d.ts +28 -30
- package/src/debug/index.js +0 -18
- package/src/debug/stats.d.ts +0 -8
- package/src/debug/stats.js +3 -3
- package/src/debug/tensor.d.ts +9 -2
- package/src/debug/tensor.js +11 -4
- package/src/debug/trace.js +2 -15
- package/src/errors/doppler-error.js +20 -0
- package/src/experimental/adapters/lora-loader.js +109 -24
- package/src/experimental/browser/quantization.d.ts +5 -0
- package/src/experimental/diffusion/index.d.ts +24 -0
- package/src/experimental/diffusion/index.js +10 -0
- package/src/experimental/distribution/descriptor-transport.d.ts +117 -0
- package/src/experimental/distribution/descriptor-transport.js +342 -0
- package/src/experimental/distribution/shard-delivery.js +179 -23
- package/src/experimental/training/README.md +11 -0
- package/src/experimental/training/artifacts.js +1 -14
- package/src/experimental/training/autograd.js +33 -8
- package/src/experimental/training/checkpoint.js +1 -14
- package/src/experimental/training/datasets/gepa-frontier.d.ts +41 -0
- package/src/experimental/training/datasets/gepa-frontier.js +220 -0
- package/src/experimental/training/datasets/index.d.ts +13 -1
- package/src/experimental/training/datasets/index.js +8 -1
- package/src/experimental/training/datasets/teacher-traces.d.ts +51 -0
- package/src/experimental/training/datasets/teacher-traces.js +279 -0
- package/src/experimental/training/datasets/text-pairs.d.ts +45 -0
- package/src/experimental/training/datasets/text-pairs.js +82 -5
- package/src/experimental/training/datasets/token-batch.d.ts +1 -1
- package/src/experimental/training/datasets/token-batch.js +2 -2
- package/src/experimental/training/distillation/dataset.js +2 -15
- package/src/experimental/training/distillation/eval.js +3 -3
- package/src/experimental/training/distillation/runtime.js +3 -16
- package/src/experimental/training/distillation/student-fixture.d.ts +5 -0
- package/src/experimental/training/distillation/student-fixture.js +296 -38
- package/src/experimental/training/export.d.ts +9 -0
- package/src/experimental/training/export.js +113 -7
- package/src/experimental/training/index.d.ts +13 -1
- package/src/experimental/training/index.js +13 -1
- package/src/experimental/training/lora-pipeline.d.ts +143 -0
- package/src/experimental/training/lora-pipeline.js +1733 -58
- package/src/experimental/training/lora.js +23 -1
- package/src/experimental/training/loss.js +1 -1
- package/src/experimental/training/objectives/base.d.ts +2 -2
- package/src/experimental/training/objectives/cross_entropy.d.ts +2 -2
- package/src/experimental/training/objectives/distill_kd.d.ts +2 -2
- package/src/experimental/training/objectives/distill_kd.js +2 -2
- package/src/experimental/training/objectives/distill_triplet.d.ts +2 -2
- package/src/experimental/training/objectives/distill_triplet.js +2 -2
- package/src/experimental/training/objectives/ul_stage1_joint.d.ts +2 -2
- package/src/experimental/training/objectives/ul_stage2_base.d.ts +2 -2
- package/src/experimental/training/objectives/ul_stage2_base.js +1 -1
- package/src/experimental/training/operator-agent-eval.d.ts +65 -0
- package/src/experimental/training/operator-agent-eval.js +326 -0
- package/src/experimental/training/operator-artifacts.js +1 -14
- package/src/experimental/training/operator-command.js +310 -51
- package/src/experimental/training/runner.d.ts +1 -0
- package/src/experimental/training/runner.js +5 -14
- package/src/experimental/training/workload-packs/distill-glm52-doppler-agent-harness-gemma270m-f16-sft-tiny.json +160 -0
- package/src/experimental/training/workload-packs/distill-glm52-doppler-code-agent-gemma270m-f16-sft-tiny.json +175 -0
- package/src/experimental/training/workload-packs/distill-glm52-doppler-code-agent-sft-tiny.json +175 -0
- package/src/experimental/training/workload-packs/distill-glm52-doppler-js-json-gemma270m-f16-sft-tiny.json +160 -0
- package/src/experimental/training/workload-packs/distill-glm52-doppler-review-gemma270m-f16-sft-tiny.json +162 -0
- package/src/experimental/training/workload-packs/distill-glm52-doppler-wgsl-gemma270m-f16-sft-tiny.json +160 -0
- package/src/experimental/training/workload-packs/lora-doppler-agent-harness-gemma270m-f16-tiny.json +137 -0
- package/src/experimental/training/workload-packs/lora-doppler-code-agent-gemma270m-f16-tiny.json +152 -0
- package/src/experimental/training/workload-packs/lora-doppler-code-agent-tiny.json +152 -0
- package/src/experimental/training/workload-packs/lora-doppler-js-json-gemma270m-f16-tiny.json +137 -0
- package/src/experimental/training/workload-packs/lora-doppler-review-gemma270m-f16-tiny.json +139 -0
- package/src/experimental/training/workload-packs/lora-doppler-wgsl-gemma270m-f16-tiny.json +137 -0
- package/src/experimental/training/workload-packs/registry.json +192 -12
- package/src/experimental/training/workloads.d.ts +37 -0
- package/src/experimental/training/workloads.js +111 -2
- package/src/formats/gguf/types.d.ts +0 -63
- package/src/formats/gguf/types.js +58 -38
- package/src/formats/litert/types.d.ts +0 -13
- package/src/formats/litert/types.js +2 -2
- package/src/formats/rdrr/distributed/index.d.ts +34 -0
- package/src/formats/rdrr/distributed/index.js +3 -0
- package/src/formats/rdrr/distributed/parsing.d.ts +6 -0
- package/src/formats/rdrr/distributed/parsing.js +23 -0
- package/src/formats/rdrr/distributed/validation.d.ts +23 -0
- package/src/formats/rdrr/distributed/validation.js +131 -0
- package/src/formats/rdrr/functional-descriptor.d.ts +73 -0
- package/src/formats/rdrr/functional-descriptor.js +230 -0
- package/src/formats/rdrr/groups.d.ts +1 -13
- package/src/formats/rdrr/groups.js +0 -34
- package/src/formats/rdrr/index.d.ts +10 -0
- package/src/formats/rdrr/index.js +10 -0
- package/src/formats/rdrr/integrity.d.ts +51 -0
- package/src/formats/rdrr/integrity.js +200 -0
- package/src/formats/rdrr/lowerings.d.ts +28 -0
- package/src/formats/rdrr/lowerings.js +112 -0
- package/src/formats/rdrr/manifest.d.ts +0 -14
- package/src/formats/rdrr/manifest.js +5 -58
- package/src/formats/rdrr/merkle.d.ts +24 -0
- package/src/formats/rdrr/merkle.js +118 -0
- package/src/formats/rdrr/parsing.js +26 -6
- package/src/formats/rdrr/slice-compiler.d.ts +29 -0
- package/src/formats/rdrr/slice-compiler.js +188 -0
- package/src/formats/rdrr/source-transform-contract.js +10 -1
- package/src/formats/rdrr/storage-descriptor.d.ts +36 -0
- package/src/formats/rdrr/storage-descriptor.js +193 -0
- package/src/formats/rdrr/tensor-config-validator.d.ts +0 -9
- package/src/formats/rdrr/tensor-config-validator.js +0 -30
- package/src/formats/rdrr/types.d.ts +111 -2
- package/src/formats/rdrr/validation.js +260 -1
- package/src/formats/tflite/types.d.ts +0 -31
- package/src/formats/tflite/types.js +5 -5
- package/src/gpu/command-recorder.d.ts +45 -23
- package/src/gpu/command-recorder.js +122 -25
- package/src/gpu/device.d.ts +17 -3
- package/src/gpu/device.js +130 -14
- package/src/gpu/kernel-runtime.d.ts +8 -15
- package/src/gpu/kernel-selection-log.d.ts +0 -8
- package/src/gpu/kernel-selection-log.js +0 -15
- package/src/gpu/kernel-tuner/cache.js +1 -8
- package/src/gpu/kernels/README.md +50 -18
- package/src/gpu/kernels/activation_static_qdq.wgsl +3 -1
- package/src/gpu/kernels/attention.d.ts +28 -19
- package/src/gpu/kernels/attention.js +458 -16
- package/src/gpu/kernels/attention_decode_contiguous_turboquant_f16kv.wgsl +0 -1
- package/src/gpu/kernels/attention_decode_contiguous_turboquant_prod_f16kv.wgsl +0 -1
- package/src/gpu/kernels/attention_decode_f16.wgsl +1 -1
- package/src/gpu/kernels/attention_decode_f16kv.wgsl +1 -1
- package/src/gpu/kernels/attention_decode_online_f16.wgsl +32 -29
- package/src/gpu/kernels/attention_decode_online_f16kv.wgsl +1 -1
- package/src/gpu/kernels/attention_decode_online_head256_f16kv.wgsl +236 -0
- package/src/gpu/kernels/attention_decode_paged_f16kv.wgsl +1 -1
- package/src/gpu/kernels/attention_decode_tiered_f16kv.wgsl +1 -1
- package/src/gpu/kernels/attention_decode_tiered_turboquant_f16kv.wgsl +0 -1
- package/src/gpu/kernels/attention_decode_tiered_turboquant_prod_f16kv.wgsl +0 -1
- package/src/gpu/kernels/attention_f16.wgsl +88 -30
- package/src/gpu/kernels/attention_f16kv.wgsl +45 -15
- package/src/gpu/kernels/attention_head512_f16.wgsl +270 -0
- package/src/gpu/kernels/attention_head512_f16kv.wgsl +271 -0
- package/src/gpu/kernels/attention_prefill_flash_head256_f16kv.wgsl +306 -0
- package/src/gpu/kernels/attention_prefill_flash_reduce.wgsl +101 -0
- package/src/gpu/kernels/attention_small_f16.wgsl +22 -20
- package/src/gpu/kernels/attention_small_f16kv.wgsl +1 -1
- package/src/gpu/kernels/attention_streaming_f16.wgsl +34 -9
- package/src/gpu/kernels/attention_streaming_f16kv.wgsl +22 -3
- package/src/gpu/kernels/backward/cross_entropy_backward.js +72 -3
- package/src/gpu/kernels/backward/cross_entropy_backward.wgsl +2 -2
- package/src/gpu/kernels/backward/index.d.ts +6 -0
- package/src/gpu/kernels/backward/layernorm_backward.wgsl +0 -1
- package/src/gpu/kernels/backward/utils.d.ts +35 -0
- package/src/gpu/kernels/bf16_to_f32.wgsl +0 -16
- package/src/gpu/kernels/bias_add_f16.wgsl +1 -1
- package/src/gpu/kernels/cast.js +74 -67
- package/src/gpu/kernels/cast_f32_to_f16.wgsl +0 -1
- package/src/gpu/kernels/codegen/patches/attention_f16.from.attention.diff +19 -215
- package/src/gpu/kernels/codegen/patches/attention_f16kv.from.attention.diff +8 -82
- package/src/gpu/kernels/codegen/patches/attention_small_f16.from.attention_small.diff +92 -59
- package/src/gpu/kernels/codegen/patches/attention_streaming_f16.from.attention_streaming.diff +12 -50
- package/src/gpu/kernels/codegen/patches/gelu_f16.from.gelu.diff +16 -33
- package/src/gpu/kernels/codegen/patches/residual_f16.from.residual.diff +15 -12
- package/src/gpu/kernels/codegen/patches/rmsnorm_f16.from.rmsnorm.diff +255 -170
- package/src/gpu/kernels/codegen/patches/rope_f16.from.rope.diff +133 -235
- package/src/gpu/kernels/codegen/patches/sample_f16.from.sample.diff +119 -90
- package/src/gpu/kernels/codegen/wgsl-variants.js +142 -8
- package/src/gpu/kernels/constants.d.ts +9 -48
- package/src/gpu/kernels/constants.js +0 -23
- package/src/gpu/kernels/conv2d.d.ts +0 -4
- package/src/gpu/kernels/conv2d_f16.wgsl +1 -1
- package/src/gpu/kernels/dequant.d.ts +0 -16
- package/src/gpu/kernels/dequant.js +1 -71
- package/src/gpu/kernels/dequant_mxfp4_expert_f16.wgsl +1 -1
- package/src/gpu/kernels/dequant_shared.wgsl +1 -55
- package/src/gpu/kernels/diffusion-gemma-sampling.d.ts +19 -0
- package/src/gpu/kernels/diffusion-gemma-sampling.js +104 -0
- package/src/gpu/kernels/diffusion_gemma_sampling.wgsl +130 -0
- package/src/gpu/kernels/dispatch.d.ts +12 -97
- package/src/gpu/kernels/dispatch.js +13 -120
- package/src/gpu/kernels/energy_eval_f16.wgsl +1 -1
- package/src/gpu/kernels/energy_quintel_grad_f16.wgsl +1 -1
- package/src/gpu/kernels/energy_quintel_reduce_f16.wgsl +1 -1
- package/src/gpu/kernels/energy_quintel_update_f16.wgsl +1 -1
- package/src/gpu/kernels/energy_update_f16.wgsl +1 -1
- package/src/gpu/kernels/fused-gate-up-gelu.js +121 -0
- package/src/gpu/kernels/fused_ffn.js +3 -4
- package/src/gpu/kernels/fused_ffn_f16.wgsl +1 -1
- package/src/gpu/kernels/fused_ffn_q4k.wgsl +0 -19
- package/src/gpu/kernels/fused_ffn_q4k_f16.wgsl +0 -18
- package/src/gpu/kernels/fused_gate_up_gelu_gated_f16.wgsl +124 -0
- package/src/gpu/kernels/fused_matmul_litert_int4_f16a.wgsl +78 -0
- package/src/gpu/kernels/fused_matmul_litert_int4_f16a_f32out.wgsl +78 -0
- package/src/gpu/kernels/fused_matmul_litert_int4_f32a_f32out.wgsl +76 -0
- package/src/gpu/kernels/fused_matmul_q4.wgsl +23 -439
- package/src/gpu/kernels/fused_matmul_q4_batched_f16.wgsl +1 -1
- package/src/gpu/kernels/fused_matmul_q4_batched_f16a.wgsl +0 -1
- package/src/gpu/kernels/fused_matmul_q4_batched_f16acc_f16a.wgsl +158 -0
- package/src/gpu/kernels/fused_matmul_q4_f16a.wgsl +0 -1
- package/src/gpu/kernels/fused_matmul_q4_multicol_f16.wgsl +0 -1
- package/src/gpu/kernels/fused_matmul_q4_multicol_f16a.wgsl +32 -33
- package/src/gpu/kernels/fused_matmul_q4_multicol_f16a_f32acc.wgsl +209 -0
- package/src/gpu/kernels/fused_matmul_q4_prefill_tiled_f16.wgsl +261 -0
- package/src/gpu/kernels/fused_matmul_q4_widetile.wgsl +164 -0
- package/src/gpu/kernels/fused_matmul_q4_widetile_f16.wgsl +198 -0
- package/src/gpu/kernels/fused_matmul_q4_widetile_f16a.wgsl +161 -0
- package/src/gpu/kernels/fused_matmul_q4_widetile_residual.wgsl +173 -0
- package/src/gpu/kernels/fused_matmul_rmsnorm.js +101 -80
- package/src/gpu/kernels/fused_matmul_rmsnorm_f16.wgsl +1 -1
- package/src/gpu/kernels/fused_matmul_w4a16_batched_f16a.wgsl +116 -0
- package/src/gpu/kernels/fused_matmul_w4a16_batched_f16a_f32out.wgsl +116 -0
- package/src/gpu/kernels/fused_matmul_w4a16_f16a.wgsl +95 -0
- package/src/gpu/kernels/fused_matmul_w4a16_f16a_f32out.wgsl +95 -0
- package/src/gpu/kernels/fused_rmsnorm_q4_widetile.wgsl +271 -0
- package/src/gpu/kernels/fused_swiglu_f16.wgsl +1 -1
- package/src/gpu/kernels/gather.d.ts +84 -8
- package/src/gpu/kernels/gather.js +353 -5
- package/src/gpu/kernels/gather_f16.wgsl +1 -1
- package/src/gpu/kernels/gather_litert_int4_vec4_f16_out.wgsl +75 -0
- package/src/gpu/kernels/gather_litert_int4_vec4_f32_out.wgsl +72 -0
- package/src/gpu/kernels/gather_split4_f16_vec4_f16_out.wgsl +103 -0
- package/src/gpu/kernels/gather_split8_f16_vec4_f16_out.wgsl +143 -0
- package/src/gpu/kernels/gather_split8_f16_vec4_f32_out.wgsl +143 -0
- package/src/gpu/kernels/gelu.wgsl +11 -7
- package/src/gpu/kernels/gelu_f16.wgsl +25 -22
- package/src/gpu/kernels/gemma4-route-expert.d.ts +35 -0
- package/src/gpu/kernels/gemma4-route-expert.js +126 -0
- package/src/gpu/kernels/gemma4_route_q4_matmul_f16a.wgsl +210 -0
- package/src/gpu/kernels/groupnorm.d.ts +0 -4
- package/src/gpu/kernels/groupnorm_apply_f16.wgsl +1 -1
- package/src/gpu/kernels/groupnorm_stats_f16.wgsl +1 -1
- package/src/gpu/kernels/index.d.ts +77 -5
- package/src/gpu/kernels/index.js +55 -4
- package/src/gpu/kernels/kv-cache-write.d.ts +22 -0
- package/src/gpu/kernels/kv-cache-write.js +126 -0
- package/src/gpu/kernels/kv_cache_write_f32_to_f16.wgsl +32 -0
- package/src/gpu/kernels/kv_quantize_turboquant.wgsl +0 -2
- package/src/gpu/kernels/kv_quantize_turboquant_prod.wgsl +0 -2
- package/src/gpu/kernels/layernorm.d.ts +0 -4
- package/src/gpu/kernels/layernorm_f16.wgsl +1 -1
- package/src/gpu/kernels/{sana_linear_attention.d.ts → linear_attention.d.ts} +5 -5
- package/src/gpu/kernels/{sana_linear_attention.js → linear_attention.js} +16 -16
- package/src/gpu/kernels/lm-head-argmax.d.ts +19 -0
- package/src/gpu/kernels/lm-head-argmax.js +218 -0
- package/src/gpu/kernels/lm_head_argmax_f16.wgsl +189 -0
- package/src/gpu/kernels/logit-merge.d.ts +0 -24
- package/src/gpu/kernels/logit-merge.js +1 -17
- package/src/gpu/kernels/matmul-dispatch.d.ts +11 -0
- package/src/gpu/kernels/matmul-dispatch.js +44 -2
- package/src/gpu/kernels/matmul-selection.d.ts +5 -5
- package/src/gpu/kernels/matmul-selection.js +172 -10
- package/src/gpu/kernels/matmul.d.ts +1 -29
- package/src/gpu/kernels/matmul.js +240 -38
- package/src/gpu/kernels/matmul_f16_tiled.wgsl +33 -34
- package/src/gpu/kernels/matmul_gemv_residual_f16.wgsl +1 -1
- package/src/gpu/kernels/matmul_gemv_subgroup.wgsl +0 -1
- package/src/gpu/kernels/matmul_gemv_subgroup_f16a.wgsl +24 -25
- package/src/gpu/kernels/modulate_f16.wgsl +1 -1
- package/src/gpu/kernels/moe.d.ts +1 -56
- package/src/gpu/kernels/moe.js +46 -33
- package/src/gpu/kernels/moe_gather_f16.wgsl +1 -1
- package/src/gpu/kernels/pipeline-cache.d.ts +8 -0
- package/src/gpu/kernels/pipeline-cache.js +24 -0
- package/src/gpu/kernels/pixel_shuffle_f16.wgsl +1 -1
- package/src/gpu/kernels/rep-penalty.d.ts +21 -0
- package/src/gpu/kernels/rep-penalty.js +17 -9
- package/src/gpu/kernels/residual.d.ts +1 -12
- package/src/gpu/kernels/residual.js +13 -1
- package/src/gpu/kernels/residual.wgsl +3 -3
- package/src/gpu/kernels/residual_f16.wgsl +6 -4
- package/src/gpu/kernels/residual_f16_vec4.wgsl +6 -6
- package/src/gpu/kernels/residual_vec4.wgsl +6 -6
- package/src/gpu/kernels/rmsnorm-pair.d.ts +39 -0
- package/src/gpu/kernels/rmsnorm-pair.js +152 -0
- package/src/gpu/kernels/rmsnorm-qk.d.ts +41 -0
- package/src/gpu/kernels/rmsnorm-qk.js +128 -0
- package/src/gpu/kernels/rmsnorm.d.ts +30 -7
- package/src/gpu/kernels/rmsnorm.js +49 -8
- package/src/gpu/kernels/rmsnorm.wgsl +15 -12
- package/src/gpu/kernels/rmsnorm_f16.wgsl +69 -39
- package/src/gpu/kernels/rmsnorm_pair.wgsl +224 -0
- package/src/gpu/kernels/rmsnorm_qk.wgsl +118 -0
- package/src/gpu/kernels/rmsnorm_qk_f16.wgsl +143 -0
- package/src/gpu/kernels/rope-qk.d.ts +41 -0
- package/src/gpu/kernels/rope-qk.js +143 -0
- package/src/gpu/kernels/rope.d.ts +1 -5
- package/src/gpu/kernels/rope.js +12 -2
- package/src/gpu/kernels/rope.wgsl +33 -19
- package/src/gpu/kernels/rope_f16.wgsl +39 -25
- package/src/gpu/kernels/rope_qk.wgsl +95 -0
- package/src/gpu/kernels/rope_qk_f16.wgsl +97 -0
- package/src/gpu/kernels/rule-registry.js +50 -10
- package/src/gpu/kernels/sample.d.ts +0 -5
- package/src/gpu/kernels/sample.js +212 -187
- package/src/gpu/kernels/sample.wgsl +7 -6
- package/src/gpu/kernels/sample_f16.wgsl +28 -30
- package/src/gpu/kernels/scatter_add_dynamic_expert_scale.wgsl +47 -0
- package/src/gpu/kernels/scatter_add_dynamic_f16.wgsl +1 -1
- package/src/gpu/kernels/scatter_add_dynamic_f16_expert_scale.wgsl +49 -0
- package/src/gpu/kernels/scatter_add_dynamic_f16_weights_expert_scale.wgsl +49 -0
- package/src/gpu/kernels/scatter_add_routes_f16_weights_expert_scale.wgsl +45 -0
- package/src/gpu/kernels/shader-cache.d.ts +13 -0
- package/src/gpu/kernels/shader-cache.js +28 -21
- package/src/gpu/kernels/silu_f16.wgsl +1 -1
- package/src/gpu/kernels/soft-embedding.d.ts +28 -0
- package/src/gpu/kernels/soft-embedding.js +447 -0
- package/src/gpu/kernels/soft_embedding_logits_f16.wgsl +107 -0
- package/src/gpu/kernels/soft_embedding_split_f16.wgsl +46 -0
- package/src/gpu/kernels/softmax.wgsl +8 -223
- package/src/gpu/kernels/split-qkv-rmsnorm-qk.d.ts +44 -0
- package/src/gpu/kernels/split-qkv-rmsnorm-qk.js +154 -0
- package/src/gpu/kernels/split-qkv-rmsnorm-rope-qk.d.ts +59 -0
- package/src/gpu/kernels/split-qkv-rmsnorm-rope-qk.js +245 -0
- package/src/gpu/kernels/split_qg_f16.wgsl +1 -1
- package/src/gpu/kernels/split_qkv_f16.wgsl +1 -1
- package/src/gpu/kernels/split_qkv_rmsnorm_qk.wgsl +146 -0
- package/src/gpu/kernels/split_qkv_rmsnorm_rope_qk.wgsl +185 -0
- package/src/gpu/kernels/split_qkv_rmsnorm_rope_qk_f16kv_cache.wgsl +187 -0
- package/src/gpu/kernels/topk_f16.wgsl +1 -1
- package/src/gpu/kernels/turboquant-codebook.js +11 -40
- package/src/gpu/kernels/uniform-utils.d.ts +18 -0
- package/src/gpu/kernels/uniform-utils.js +2 -5
- package/src/gpu/kernels/upsample2d.d.ts +0 -4
- package/src/gpu/kernels/upsample2d_f16.wgsl +1 -1
- package/src/gpu/kernels/utils.d.ts +43 -0
- package/src/gpu/kernels/utils.js +81 -11
- package/src/gpu/perf-guards.d.ts +0 -12
- package/src/gpu/perf-guards.js +0 -62
- package/src/gpu/profiler.d.ts +0 -21
- package/src/gpu/profiler.js +0 -34
- package/src/gpu/readback-utils.js +10 -2
- package/src/gpu/submit-tracker.d.ts +0 -57
- package/src/gpu/submit-tracker.js +3 -62
- package/src/gpu/tensor.d.ts +0 -23
- package/src/gpu/tensor.js +0 -41
- package/src/gpu/uniform-cache.d.ts +10 -12
- package/src/gpu/uniform-cache.js +25 -30
- package/src/gpu/weight-buffer.d.ts +76 -7
- package/src/gpu/weight-buffer.js +85 -1
- package/src/index.d.ts +1 -0
- package/src/index.js +1 -0
- package/src/inference/browser-harness-diffusion-energy-suites.js +269 -1
- package/src/inference/browser-harness-model-helpers.d.ts +0 -9
- package/src/inference/browser-harness-model-helpers.js +22 -13
- package/src/inference/browser-harness-runtime-helpers.d.ts +0 -5
- package/src/inference/browser-harness-runtime-helpers.js +2 -2
- package/src/inference/browser-harness-suite-helpers.d.ts +13 -1
- package/src/inference/browser-harness-suite-helpers.js +525 -3
- package/src/inference/browser-harness-text-helpers.d.ts +89 -0
- package/src/inference/browser-harness-text-helpers.js +459 -25
- package/src/inference/browser-harness.d.ts +7 -28
- package/src/inference/browser-harness.js +556 -79
- package/src/inference/decode-ring.d.ts +1 -0
- package/src/inference/decode-ring.js +7 -5
- package/src/inference/kv-cache/base.d.ts +31 -0
- package/src/inference/kv-cache/base.js +125 -2
- package/src/inference/kv-cache/mixed-geometry.d.ts +11 -0
- package/src/inference/kv-cache/mixed-geometry.js +121 -4
- package/src/inference/kv-cache/quantized.js +2 -0
- package/src/inference/kv-cache/sliding-window.d.ts +14 -0
- package/src/inference/kv-cache/sliding-window.js +88 -0
- package/src/inference/kv-cache/tiered.js +7 -0
- package/src/inference/kv-cache/types.d.ts +2 -0
- package/src/inference/model-lanes/execution-lane-resolver.d.ts +61 -0
- package/src/inference/model-lanes/execution-lane-resolver.js +126 -0
- package/src/inference/moe-router.d.ts +14 -1
- package/src/inference/moe-router.js +24 -1
- package/src/inference/multi-model-network.d.ts +1 -1
- package/src/inference/multi-model-network.js +1 -2
- package/src/inference/network-evolution.d.ts +0 -12
- package/src/inference/network-evolution.js +2 -2
- package/src/inference/pipelines/audio/gemma4.js +52 -1
- package/src/inference/pipelines/context.js +17 -0
- package/src/inference/pipelines/diffusion/init.js +2 -2
- package/src/inference/pipelines/diffusion/pipeline.js +41 -96
- package/src/inference/pipelines/diffusion/text-encoder-gpu.d.ts +0 -5
- package/src/inference/pipelines/diffusion/text-encoder-gpu.js +0 -283
- package/src/inference/pipelines/diffusion/vae.js +4 -4
- package/src/inference/pipelines/diffusion-gemma/config.d.ts +27 -0
- package/src/inference/pipelines/diffusion-gemma/config.js +149 -0
- package/src/inference/pipelines/diffusion-gemma/index.d.ts +31 -0
- package/src/inference/pipelines/diffusion-gemma/index.js +12 -0
- package/src/inference/pipelines/diffusion-gemma/pipeline.d.ts +69 -0
- package/src/inference/pipelines/diffusion-gemma/pipeline.js +380 -0
- package/src/inference/pipelines/diffusion-gemma/sampling.d.ts +128 -0
- package/src/inference/pipelines/diffusion-gemma/sampling.js +395 -0
- package/src/inference/pipelines/energy/quintel.d.ts +15 -0
- package/src/inference/pipelines/structured/index.d.ts +2 -0
- package/src/inference/pipelines/structured/index.js +2 -0
- package/src/inference/pipelines/structured/json-grammar-mask.js +143 -0
- package/src/inference/pipelines/structured/tsv-grammar-mask.js +166 -0
- package/src/inference/pipelines/text/attention/attn-config.d.ts +0 -5
- package/src/inference/pipelines/text/attention/attn-config.js +1 -1
- package/src/inference/pipelines/text/attention/dispatch-params.d.ts +28 -4
- package/src/inference/pipelines/text/attention/dispatch-params.js +176 -2
- package/src/inference/pipelines/text/attention/projections.d.ts +73 -2
- package/src/inference/pipelines/text/attention/projections.js +398 -71
- package/src/inference/pipelines/text/attention/quant-options.d.ts +38 -0
- package/src/inference/pipelines/text/attention/record.js +542 -56
- package/src/inference/pipelines/text/attention/rmsnorm-fusion-gate.d.ts +8 -0
- package/src/inference/pipelines/text/attention/rmsnorm-fusion-gate.js +19 -0
- package/src/inference/pipelines/text/attention/run.js +343 -50
- package/src/inference/pipelines/text/attention/types.d.ts +7 -1
- package/src/inference/pipelines/text/chat-format.d.ts +0 -10
- package/src/inference/pipelines/text/chat-format.js +1 -8
- package/src/inference/pipelines/text/config.d.ts +25 -26
- package/src/inference/pipelines/text/config.js +304 -25
- package/src/inference/pipelines/text/debug-utils/utils.js +2 -23
- package/src/inference/pipelines/text/drift-policy.js +1 -1
- package/src/inference/pipelines/text/dtype-contract.d.ts +1 -1
- package/src/inference/pipelines/text/dtype-contract.js +4 -1
- package/src/inference/pipelines/text/embed.d.ts +25 -29
- package/src/inference/pipelines/text/embed.js +16 -45
- package/src/inference/pipelines/text/execution-plan.d.ts +13 -3
- package/src/inference/pipelines/text/execution-plan.js +21 -2
- package/src/inference/pipelines/text/execution-runtime-builders.d.ts +4 -4
- package/src/inference/pipelines/text/execution-runtime-builders.js +44 -39
- package/src/inference/pipelines/text/execution-v1.d.ts +20 -0
- package/src/inference/pipelines/text/execution-v1.js +461 -158
- package/src/inference/pipelines/text/ffn/dense.js +214 -6
- package/src/inference/pipelines/text/ffn/moe.d.ts +4 -1
- package/src/inference/pipelines/text/ffn/moe.js +9 -2
- package/src/inference/pipelines/text/ffn/sandwich.d.ts +3 -1
- package/src/inference/pipelines/text/ffn/sandwich.js +237 -15
- package/src/inference/pipelines/text/ffn/standard.d.ts +3 -1
- package/src/inference/pipelines/text/ffn/standard.js +84 -17
- package/src/inference/pipelines/text/generator-decode-policy.d.ts +81 -0
- package/src/inference/pipelines/text/generator-decode-policy.js +233 -0
- package/src/inference/pipelines/text/generator-helpers.d.ts +0 -2
- package/src/inference/pipelines/text/generator-helpers.js +9 -7
- package/src/inference/pipelines/text/generator-prefill-helpers.d.ts +84 -0
- package/src/inference/pipelines/text/generator-prefill-helpers.js +359 -0
- package/src/inference/pipelines/text/generator-runtime.d.ts +1 -7
- package/src/inference/pipelines/text/generator-runtime.js +8 -6
- package/src/inference/pipelines/text/generator-steps.d.ts +57 -7
- package/src/inference/pipelines/text/generator-steps.js +275 -79
- package/src/inference/pipelines/text/generator.d.ts +38 -5
- package/src/inference/pipelines/text/generator.js +1259 -496
- package/src/inference/pipelines/text/init-chat-templates.js +86 -0
- package/src/inference/pipelines/text/init-qkv-fusion.js +201 -0
- package/src/inference/pipelines/text/init-rope.js +408 -0
- package/src/inference/pipelines/text/init.d.ts +13 -3
- package/src/inference/pipelines/text/init.js +213 -725
- package/src/inference/pipelines/text/kernel-trace.d.ts +0 -33
- package/src/inference/pipelines/text/kernel-trace.js +0 -33
- package/src/inference/pipelines/text/layer-plan-gpu.js +590 -0
- package/src/inference/pipelines/text/layer-plan.d.ts +0 -5
- package/src/inference/pipelines/text/layer-plan.js +1 -1
- package/src/inference/pipelines/text/layer.d.ts +70 -8
- package/src/inference/pipelines/text/layer.js +373 -558
- package/src/inference/pipelines/text/linear-attention.d.ts +2 -1
- package/src/inference/pipelines/text/linear-attention.js +6 -4
- package/src/inference/pipelines/text/load-errors.d.ts +21 -0
- package/src/inference/pipelines/text/load-errors.js +59 -0
- package/src/inference/pipelines/text/logits/cpu.js +2 -20
- package/src/inference/pipelines/text/logits/gpu.d.ts +49 -2
- package/src/inference/pipelines/text/logits/gpu.js +419 -66
- package/src/inference/pipelines/text/logits/index.d.ts +3 -3
- package/src/inference/pipelines/text/logits/index.js +96 -42
- package/src/inference/pipelines/text/logits/types.d.ts +2 -2
- package/src/inference/pipelines/text/lora-apply.js +17 -0
- package/src/inference/pipelines/text/model-load.d.ts +5 -4
- package/src/inference/pipelines/text/model-load.js +74 -17
- package/src/inference/pipelines/text/moe-gpu.d.ts +13 -0
- package/src/inference/pipelines/text/moe-gpu.js +445 -16
- package/src/inference/pipelines/text/moe-impl.d.ts +13 -3
- package/src/inference/pipelines/text/moe-shape-validator.js +13 -1
- package/src/inference/pipelines/text/ops.d.ts +42 -4
- package/src/inference/pipelines/text/ops.js +20 -1
- package/src/inference/pipelines/text/per-layer-inputs.js +36 -6
- package/src/inference/pipelines/text/probes.d.ts +3 -6
- package/src/inference/pipelines/text/probes.js +119 -31
- package/src/inference/pipelines/text/sampling-config.d.ts +4 -10
- package/src/inference/pipelines/text/sampling-config.js +103 -64
- package/src/inference/pipelines/text/sampling.d.ts +0 -5
- package/src/inference/pipelines/text/sampling.js +153 -9
- package/src/inference/pipelines/text/stage-names.js +3 -0
- package/src/inference/pipelines/text/state.js +19 -0
- package/src/inference/pipelines/text/tsir-fixture-writer.d.ts +57 -0
- package/src/inference/pipelines/text/tsir-fixture-writer.js +408 -0
- package/src/inference/pipelines/text/types.d.ts +92 -81
- package/src/inference/pipelines/text/weights.d.ts +4 -52
- package/src/inference/pipelines/text/weights.js +6 -56
- package/src/inference/pipelines/text.d.ts +33 -1
- package/src/inference/pipelines/text.js +375 -55
- package/src/inference/pipelines/vision/gemma4.js +20 -9
- package/src/inference/pipelines/vision/index.js +0 -107
- package/src/inference/runtime-model.d.ts +0 -7
- package/src/inference/runtime-model.js +0 -23
- package/src/inference/test-harness.d.ts +6 -22
- package/src/inference/test-harness.js +97 -23
- package/src/inference/tokenizer.d.ts +30 -0
- package/src/inference/tokenizer.js +341 -174
- package/src/inference/tokenizers/bundled.js +71 -49
- package/src/inference/tokenizers/sentencepiece.js +65 -15
- package/src/inference/tokenizers/types.d.ts +0 -21
- package/src/loader/doppler-loader.d.ts +14 -0
- package/src/loader/doppler-loader.js +378 -25
- package/src/loader/embedding-limit-preflight.d.ts +31 -0
- package/src/loader/embedding-limit-preflight.js +216 -0
- package/src/loader/embedding-loader.d.ts +8 -2
- package/src/loader/embedding-loader.js +276 -11
- package/src/loader/experts/expert-loader.d.ts +4 -1
- package/src/loader/experts/expert-loader.js +174 -5
- package/src/loader/final-weights-loader.d.ts +5 -0
- package/src/loader/final-weights-loader.js +167 -4
- package/src/loader/layer-loader.d.ts +2 -0
- package/src/loader/layer-loader.js +74 -6
- package/src/loader/loader-types.d.ts +61 -2
- package/src/loader/manifest-config.d.ts +8 -0
- package/src/loader/manifest-config.js +25 -1
- package/src/loader/memory-monitor.d.ts +17 -5
- package/src/loader/quantization-constants.d.ts +6 -0
- package/src/loader/shard-resolver.d.ts +1 -0
- package/src/loader/shard-resolver.js +52 -4
- package/src/loader/tensors/functional-descriptor-loader.d.ts +9 -0
- package/src/loader/tensors/functional-descriptor-loader.js +377 -0
- package/src/loader/tensors/source-transform.js +6 -1
- package/src/loader/tensors/tensor-loader.d.ts +36 -8
- package/src/loader/tensors/tensor-loader.js +521 -22
- package/src/loader/tensors/tensor-reader.d.ts +4 -1
- package/src/loader/tensors/tensor-reader.js +11 -4
- package/src/loader/weights.d.ts +3 -1
- package/src/memory/address-table.d.ts +1 -9
- package/src/memory/buffer-pool.d.ts +0 -16
- package/src/memory/buffer-pool.js +13 -42
- package/src/memory/heap-manager.d.ts +18 -43
- package/src/memory/unified-detect.d.ts +19 -34
- package/src/models/diffusiongemma.d.ts +7 -0
- package/src/models/diffusiongemma.js +25 -0
- package/src/models/embeddinggemma.d.ts +7 -0
- package/src/models/embeddinggemma.js +24 -0
- package/src/models/family.d.ts +24 -0
- package/src/models/family.js +25 -0
- package/src/models/gemma3.d.ts +7 -0
- package/src/models/gemma3.js +33 -0
- package/src/models/gemma4.d.ts +7 -0
- package/src/models/gemma4.js +57 -0
- package/src/models/qwen3.d.ts +7 -0
- package/src/models/qwen3.js +36 -0
- package/src/rules/converter/tokenizer.rules.json +5 -0
- package/src/rules/execution-rules-contract-check.js +162 -10
- package/src/rules/inference/capability-transforms.rules.json +111 -0
- package/src/rules/inference/config.rules.json +1 -0
- package/src/rules/inference/dtype.rules.json +7 -0
- package/src/rules/inference/execution.rules.json +117 -16
- package/src/rules/inference/ffn.rules.json +24 -2
- package/src/rules/kernels/gather.rules.json +81 -0
- package/src/rules/kernels/matmul.rules.json +19 -0
- package/src/rules/kernels/moe.rules.json +3 -0
- package/src/rules/kernels/rep-penalty.rules.json +6 -0
- package/src/rules/kernels/rmsnorm-qk.rules.json +6 -0
- package/src/rules/kernels/rmsnorm.rules.json +42 -7
- package/src/rules/kernels/rope-qk.rules.json +6 -0
- package/src/rules/loader/tensor-loader.rules.json +17 -0
- package/src/rules/rule-registry.js +59 -75
- package/src/storage/artifact-storage-context.d.ts +12 -3
- package/src/storage/artifact-storage-context.js +241 -26
- package/src/storage/backends/opfs-store.js +138 -88
- package/src/storage/blake3.js +9 -1
- package/src/storage/export.js +1 -5
- package/src/storage/normalize-model-id.js +3 -0
- package/src/storage/preflight.js +11 -0
- package/src/storage/quickstart-downloader.js +44 -67
- package/src/storage/reports.js +35 -11
- package/src/storage/shard-manager.d.ts +8 -2
- package/src/storage/shard-manager.js +176 -15
- package/src/storage/source-artifact-store.d.ts +0 -2
- package/src/storage/source-artifact-store.js +3 -9
- package/src/sw.js +0 -3
- package/src/tooling/browser-command-runner.d.ts +0 -3
- package/src/tooling/browser-command-runner.js +31 -4
- package/src/tooling/command-api-constants.js +1 -1
- package/src/tooling/command-api-family-normalizers.d.ts +1 -0
- package/src/tooling/command-api-family-normalizers.js +53 -9
- package/src/tooling/command-api-helpers.d.ts +0 -2
- package/src/tooling/command-api-helpers.js +8 -3
- package/src/tooling/command-api.d.ts +22 -1
- package/src/tooling/command-api.js +11 -1
- package/src/tooling/command-envelope.d.ts +0 -1
- package/src/tooling/command-envelope.js +1 -1
- package/src/tooling/command-runner-shared.d.ts +14 -0
- package/src/tooling/command-runner-shared.js +2 -1
- package/src/tooling/conversion-config-materializer.d.ts +0 -10
- package/src/tooling/conversion-config-materializer.js +1 -1
- package/src/tooling/diagnose-runner.js +3 -4
- package/src/tooling/hf-registry-utils.d.ts +2 -6
- package/src/tooling/hf-registry-utils.js +90 -19
- package/src/tooling/kernel-path-builder/index.js +1 -4
- package/src/tooling/kernel-path-builder/runtime-overlay.js +1 -3
- package/src/tooling/lean-execution-contract-runner.d.ts +0 -10
- package/src/tooling/lean-execution-contract-runner.js +2 -2
- package/src/tooling/litert-package-runtime.d.ts +1 -1
- package/src/tooling/litert-package-runtime.js +571 -37
- package/src/tooling/node-browser-command-runner.d.ts +15 -0
- package/src/tooling/node-browser-command-runner.js +107 -17
- package/src/tooling/node-command-runner.js +43 -2
- package/src/tooling/node-convert-worker-pool.js +15 -0
- package/src/tooling/node-convert-worker.js +11 -1
- package/src/tooling/node-converter.js +102 -1
- package/src/tooling/node-source-runtime.js +1 -10
- package/src/tooling/node-webgpu.d.ts +1 -0
- package/src/tooling/node-webgpu.js +23 -11
- package/src/tooling/opfs-cache.js +1 -4
- package/src/tooling/precision-replay-math.js +174 -0
- package/src/tooling/program-bundle-host.d.ts +16 -0
- package/src/tooling/program-bundle-host.js +20 -0
- package/src/tooling/program-bundle-parity.d.ts +28 -0
- package/src/tooling/program-bundle-parity.js +210 -0
- package/src/tooling/program-bundle.d.ts +112 -0
- package/src/tooling/program-bundle.js +1119 -0
- package/src/tooling/rdrr-integrity-refresh.d.ts +50 -0
- package/src/tooling/rdrr-integrity-refresh.js +271 -0
- package/src/tooling/reference-verify.js +191 -0
- package/src/tooling/source-artifact-adapter.d.ts +0 -20
- package/src/tooling/source-artifact-adapter.js +10 -5
- package/src/tooling/source-package-profiles.d.ts +18 -1
- package/src/tooling/source-package-profiles.js +110 -8
- package/src/tooling/source-runtime-bundle.d.ts +3 -0
- package/src/tooling/source-runtime-bundle.js +30 -22
- package/src/tooling/source-runtime-converter-config.d.ts +0 -5
- package/src/tooling/source-runtime-converter-config.js +25 -12
- package/src/tooling/source-runtime-materializer.js +1 -7
- package/src/tooling-experimental-exports.shared.d.ts +10 -0
- package/src/tooling-experimental-exports.shared.js +10 -0
- package/src/tooling-exports/device.d.ts +11 -0
- package/src/tooling-exports/device.js +14 -0
- package/src/tooling-exports/manifest.d.ts +8 -0
- package/src/tooling-exports/manifest.js +11 -0
- package/src/tooling-exports/storage.d.ts +25 -0
- package/src/tooling-exports/storage.js +29 -0
- package/src/tooling-exports/structured.d.ts +1 -0
- package/src/tooling-exports/structured.js +5 -0
- package/src/tooling-exports.d.ts +23 -0
- package/src/tooling-exports.js +17 -0
- package/src/tooling-exports.shared.d.ts +27 -30
- package/src/tooling-exports.shared.js +28 -29
- package/src/types/index.d.ts +0 -3
- package/src/types/index.js +0 -3
- package/src/utils/array-buffer.d.ts +4 -0
- package/src/utils/array-buffer.js +9 -0
- package/src/utils/canonical-hash.d.ts +7 -0
- package/src/utils/canonical-hash.js +74 -0
- package/src/utils/clone-json.d.ts +1 -0
- package/src/utils/clone-json.js +7 -0
- package/src/utils/encode-utf8.js +5 -0
- package/src/utils/index.d.ts +4 -1
- package/src/utils/index.js +4 -1
- package/src/utils/runtime-env.js +7 -0
- package/src/utils/sha256.d.ts +1 -0
- package/src/utils/sha256.js +18 -4
- package/src/utils/stable-sort-object.d.ts +1 -0
- package/src/utils/stable-sort-object.js +13 -0
- package/src/version.js +1 -1
- package/tools/convert-safetensors-node.js +1 -2
- package/src/bootstrap.d.ts +0 -1
- package/src/config/conversion/sana/sana-sprint-0.6b-f16.json +0 -247
- package/src/config/kernels/kernel-ref.d.ts +0 -17
- package/src/config/kernels/kernel-ref.js +0 -75
- package/src/config/runtime/kernels/dequant-f16-q4k.json +0 -13
- package/src/config/runtime/kernels/dequant-f32-q4k.json +0 -13
- package/src/config/runtime/kernels/embeddinggemma-q4k-dequant-f32a.json +0 -37
- package/src/config/runtime/kernels/fused-q4k.json +0 -13
- package/src/config/runtime/kernels/gemma2-q4k-dequant-f16a.json +0 -33
- package/src/config/runtime/kernels/gemma2-q4k-dequant-f32a-nosubgroups.json +0 -33
- package/src/config/runtime/kernels/gemma2-q4k-fused-f32a.json +0 -33
- package/src/config/runtime/kernels/safe-q4k.json +0 -13
- package/src/config/runtime/overlays/debug-f32a.json +0 -3
- package/src/config/runtime/overlays/trace-config.json +0 -28
- package/src/config/runtime/profiles/gemma4-e2b-throughput.json +0 -39
- package/src/debug/perf.d.ts +0 -33
- package/src/debug/perf.js +0 -51
- package/src/errors/index.d.ts +0 -1
- package/src/errors/index.js +0 -1
- package/src/formats/gguf/index.d.ts +0 -8
- package/src/formats/gguf/index.js +0 -4
- package/src/formats/litert/index.d.ts +0 -1
- package/src/formats/litert/index.js +0 -1
- package/src/formats/safetensors/index.d.ts +0 -8
- package/src/formats/safetensors/index.js +0 -4
- package/src/formats/tflite/index.d.ts +0 -1
- package/src/formats/tflite/index.js +0 -1
- package/src/gpu/kernels/attention_decode_optimized.wgsl +0 -445
- package/src/gpu/kernels/dequant_mxfp4_vec4.wgsl +0 -116
- package/src/gpu/kernels/dequant_q8_0.wgsl +0 -98
- package/src/gpu/kernels/dequant_shared_vec4.wgsl +0 -160
- package/src/gpu/kernels/moe_gather_vec4.wgsl +0 -74
- package/src/inference/functiongemma.d.ts +0 -15
- package/src/inference/functiongemma.js +0 -1
- package/src/inference/pipelines/diffusion/sana-transformer.d.ts +0 -53
- package/src/inference/pipelines/diffusion/sana-transformer.js +0 -738
- package/src/rules/kernels/kernel-validator.d.ts +0 -24
- package/src/rules/kernels/kernel-validator.js +0 -160
- package/src/sw.d.ts +0 -1
- package/src/tooling/bench-runner.js +0 -64
- package/src/tooling/distill-runner.js +0 -56
- package/src/tooling/lora-runner.js +0 -37
- package/src/tooling/verify-runner.js +0 -63
- package/src/types/gpu.d.ts +0 -185
- package/src/types/gpu.js +0 -5
- package/src/types/inference.d.ts +0 -197
- package/src/types/inference.js +0 -5
- package/src/types/model.d.ts +0 -130
- package/src/types/model.js +0 -5
- package/src/utils/format-bytes.js +0 -20
- /package/src/config/{runtime/diagnostics → diagnostics}/drift-policies.json +0 -0
- /package/src/gpu/kernels/{sana_linear_attention_apply.wgsl → linear_attention_apply.wgsl} +0 -0
- /package/src/gpu/kernels/{sana_linear_attention_apply_f16.wgsl → linear_attention_apply_f16.wgsl} +0 -0
- /package/src/gpu/kernels/{sana_linear_attention_summary.wgsl → linear_attention_summary.wgsl} +0 -0
- /package/src/gpu/kernels/{sana_linear_attention_summary_f16.wgsl → linear_attention_summary_f16.wgsl} +0 -0
- /package/src/rules/kernels/{sana-linear-attention.rules.json → linear-attention.rules.json} +0 -0
|
@@ -17,6 +17,20 @@ import { isBufferActive, releaseBuffer } from '../../memory/buffer-pool.js';
|
|
|
17
17
|
|
|
18
18
|
|
|
19
19
|
export async function preloadShardsForExpert(ctx, layerIdx, expertIdx, options) {
|
|
20
|
+
const packedShardIndices = getPackedExpertShardIndices(ctx, layerIdx);
|
|
21
|
+
if (packedShardIndices.length > 0) {
|
|
22
|
+
for (const shardIndex of packedShardIndices) {
|
|
23
|
+
if (!ctx.shardCache.has(shardIndex)) {
|
|
24
|
+
await ctx.loadShard(shardIndex, options);
|
|
25
|
+
}
|
|
26
|
+
}
|
|
27
|
+
return;
|
|
28
|
+
}
|
|
29
|
+
|
|
30
|
+
if (isPackedExpertFormat(ctx.manifest?.moeConfig?.expertFormat)) {
|
|
31
|
+
return;
|
|
32
|
+
}
|
|
33
|
+
|
|
20
34
|
// Get required shards from manifest mapping
|
|
21
35
|
const shardIndices = getShardsForExpert(layerIdx, expertIdx, ctx.manifest);
|
|
22
36
|
if (shardIndices.length === 0) {
|
|
@@ -87,7 +101,7 @@ export async function loadExpert(ctx, layerIdx, expertIdx) {
|
|
|
87
101
|
}
|
|
88
102
|
}
|
|
89
103
|
|
|
90
|
-
// Fall back to simple map for non-cached
|
|
104
|
+
// Fall back to simple map for non-cached packed experts.
|
|
91
105
|
const key = `layer_${layerIdx}_expert_${expertIdx}`;
|
|
92
106
|
if (ctx.experts.has(key)) {
|
|
93
107
|
return ctx.experts.get(key);
|
|
@@ -99,16 +113,19 @@ export async function loadExpert(ctx, layerIdx, expertIdx) {
|
|
|
99
113
|
await preloadShardsForExpert(ctx, layerIdx, expertIdx);
|
|
100
114
|
|
|
101
115
|
// Get tensor names from manifest if available (for logging/debugging)
|
|
102
|
-
const
|
|
116
|
+
const expertFormat = resolveExpertFormat(ctx);
|
|
117
|
+
const tensorNames = getExpertTensorNames(ctx, layerIdx, expertIdx, expertFormat);
|
|
103
118
|
if (tensorNames.length > 0) {
|
|
104
119
|
debugTrace.loader(`Expert ${layerIdx}_${expertIdx} tensors: ${tensorNames.length}`);
|
|
105
120
|
}
|
|
106
121
|
|
|
107
|
-
const expertFormat = resolveExpertFormat(ctx);
|
|
108
122
|
let weights;
|
|
109
123
|
if (expertFormat === 'gpt-oss') {
|
|
110
124
|
weights = await loadGptOssStyleExpert(ctx, layerIdx, expertIdx);
|
|
111
125
|
assertGptOssWeights(weights, layerIdx, expertIdx);
|
|
126
|
+
} else if (expertFormat === 'gemma4') {
|
|
127
|
+
weights = await loadGemma4StyleExpert(ctx, layerIdx, expertIdx);
|
|
128
|
+
assertGemma4Weights(weights, layerIdx, expertIdx);
|
|
112
129
|
} else {
|
|
113
130
|
weights = await loadMixtralStyleExpert(ctx, layerIdx, expertIdx);
|
|
114
131
|
assertMixtralWeights(weights, layerIdx, expertIdx);
|
|
@@ -125,7 +142,7 @@ export async function loadExpert(ctx, layerIdx, expertIdx) {
|
|
|
125
142
|
const sizeBytes = calculateExpertSize(weights);
|
|
126
143
|
ctx.expertCache.put(layerIdx, expertIdx, weights, sizeBytes);
|
|
127
144
|
} else {
|
|
128
|
-
//
|
|
145
|
+
// Packed expert formats use the simple map (shared across experts).
|
|
129
146
|
ctx.experts.set(key, weights);
|
|
130
147
|
}
|
|
131
148
|
|
|
@@ -151,6 +168,10 @@ async function loadMixtralStyleExpert(ctx, layerIdx, expertIdx) {
|
|
|
151
168
|
};
|
|
152
169
|
}
|
|
153
170
|
|
|
171
|
+
function isPackedExpertFormat(expertFormat) {
|
|
172
|
+
return expertFormat === 'gpt-oss' || expertFormat === 'gemma4';
|
|
173
|
+
}
|
|
174
|
+
|
|
154
175
|
function resolveExpertFormat(ctx) {
|
|
155
176
|
const manifest = ctx.manifest ?? {};
|
|
156
177
|
const moeConfig = manifest.moeConfig ?? null;
|
|
@@ -163,7 +184,7 @@ function resolveExpertFormat(ctx) {
|
|
|
163
184
|
}
|
|
164
185
|
|
|
165
186
|
const expertFormat = moeConfig.expertFormat;
|
|
166
|
-
if (expertFormat === 'gpt-oss' || expertFormat === 'mixtral') {
|
|
187
|
+
if (expertFormat === 'gpt-oss' || expertFormat === 'mixtral' || expertFormat === 'gemma4') {
|
|
167
188
|
return expertFormat;
|
|
168
189
|
}
|
|
169
190
|
if (expertFormat == null) {
|
|
@@ -175,6 +196,89 @@ function resolveExpertFormat(ctx) {
|
|
|
175
196
|
throw new Error(`[MoE] Manifest "${modelId}" has invalid expertFormat "${expertFormat}".`);
|
|
176
197
|
}
|
|
177
198
|
|
|
199
|
+
function getShardIndicesForLocation(location) {
|
|
200
|
+
if (!location) return [];
|
|
201
|
+
if (Array.isArray(location.spans) && location.spans.length > 0) {
|
|
202
|
+
return Array.from(new Set(location.spans.map((span) => span?.shardIndex).filter(Number.isInteger)));
|
|
203
|
+
}
|
|
204
|
+
return Number.isInteger(location.shardIndex) ? [location.shardIndex] : [];
|
|
205
|
+
}
|
|
206
|
+
|
|
207
|
+
function getExistingTensorNames(ctx, candidates) {
|
|
208
|
+
const tensorLocations = ctx.tensorLocations;
|
|
209
|
+
if (!tensorLocations || typeof tensorLocations.get !== 'function') {
|
|
210
|
+
return [];
|
|
211
|
+
}
|
|
212
|
+
const names = [];
|
|
213
|
+
for (const candidateGroup of candidates) {
|
|
214
|
+
const found = candidateGroup.find((name) => tensorLocations.has(name));
|
|
215
|
+
if (found) {
|
|
216
|
+
names.push(found);
|
|
217
|
+
}
|
|
218
|
+
}
|
|
219
|
+
return names;
|
|
220
|
+
}
|
|
221
|
+
|
|
222
|
+
function getGemma4ExpertTensorCandidates(layerIdx) {
|
|
223
|
+
return [
|
|
224
|
+
[
|
|
225
|
+
`model.decoder.layers.${layerIdx}.experts.gate_up_proj`,
|
|
226
|
+
`model.encoder.language_model.layers.${layerIdx}.experts.gate_up_proj`,
|
|
227
|
+
`model.layers.${layerIdx}.experts.gate_up_proj`,
|
|
228
|
+
],
|
|
229
|
+
[
|
|
230
|
+
`model.decoder.layers.${layerIdx}.experts.down_proj`,
|
|
231
|
+
`model.encoder.language_model.layers.${layerIdx}.experts.down_proj`,
|
|
232
|
+
`model.layers.${layerIdx}.experts.down_proj`,
|
|
233
|
+
],
|
|
234
|
+
];
|
|
235
|
+
}
|
|
236
|
+
|
|
237
|
+
function getGptOssExpertTensorCandidates(layerIdx) {
|
|
238
|
+
const prefix = `model.layers.${layerIdx}.mlp.experts`;
|
|
239
|
+
return [
|
|
240
|
+
[`${prefix}.gate_up_proj_blocks`],
|
|
241
|
+
[`${prefix}.gate_up_proj_scales`],
|
|
242
|
+
[`${prefix}.gate_up_proj_bias`],
|
|
243
|
+
[`${prefix}.down_proj_blocks`],
|
|
244
|
+
[`${prefix}.down_proj_scales`],
|
|
245
|
+
[`${prefix}.down_proj_bias`],
|
|
246
|
+
];
|
|
247
|
+
}
|
|
248
|
+
|
|
249
|
+
function getPackedExpertTensorNames(ctx, layerIdx, expertFormat = ctx.manifest?.moeConfig?.expertFormat) {
|
|
250
|
+
if (expertFormat === 'gemma4') {
|
|
251
|
+
return getExistingTensorNames(ctx, getGemma4ExpertTensorCandidates(layerIdx));
|
|
252
|
+
}
|
|
253
|
+
if (expertFormat === 'gpt-oss') {
|
|
254
|
+
return getExistingTensorNames(ctx, getGptOssExpertTensorCandidates(layerIdx));
|
|
255
|
+
}
|
|
256
|
+
return [];
|
|
257
|
+
}
|
|
258
|
+
|
|
259
|
+
function getPackedExpertShardIndices(ctx, layerIdx) {
|
|
260
|
+
const tensorNames = getPackedExpertTensorNames(ctx, layerIdx);
|
|
261
|
+
const shardIndices = new Set();
|
|
262
|
+
for (const name of tensorNames) {
|
|
263
|
+
const location = ctx.tensorLocations?.get?.(name);
|
|
264
|
+
for (const shardIndex of getShardIndicesForLocation(location)) {
|
|
265
|
+
shardIndices.add(shardIndex);
|
|
266
|
+
}
|
|
267
|
+
}
|
|
268
|
+
return Array.from(shardIndices);
|
|
269
|
+
}
|
|
270
|
+
|
|
271
|
+
function getExpertTensorNames(ctx, layerIdx, expertIdx, expertFormat) {
|
|
272
|
+
const packedTensorNames = getPackedExpertTensorNames(ctx, layerIdx, expertFormat);
|
|
273
|
+
if (packedTensorNames.length > 0) {
|
|
274
|
+
return packedTensorNames;
|
|
275
|
+
}
|
|
276
|
+
if (isPackedExpertFormat(expertFormat) && !ctx.manifest?.groups?.[`layer.${layerIdx}.expert.${expertIdx}`]) {
|
|
277
|
+
return [];
|
|
278
|
+
}
|
|
279
|
+
return getTensorsForExpert(layerIdx, expertIdx, ctx.manifest);
|
|
280
|
+
}
|
|
281
|
+
|
|
178
282
|
function resolveGptOssNumExperts(ctx) {
|
|
179
283
|
const manifest = ctx.manifest ?? {};
|
|
180
284
|
const numExperts = manifest.moeConfig?.numExperts ?? null;
|
|
@@ -187,6 +291,34 @@ function resolveGptOssNumExperts(ctx) {
|
|
|
187
291
|
return numExperts;
|
|
188
292
|
}
|
|
189
293
|
|
|
294
|
+
function resolveGemma4NumExperts(ctx) {
|
|
295
|
+
const manifest = ctx.manifest ?? {};
|
|
296
|
+
const numExperts = manifest.moeConfig?.numExperts ?? null;
|
|
297
|
+
if (numExperts == null) {
|
|
298
|
+
const modelId = manifest.modelId ?? 'unknown';
|
|
299
|
+
throw new Error(`[MoE] Gemma-style manifest "${modelId}" missing moeConfig.numExperts`);
|
|
300
|
+
}
|
|
301
|
+
return numExperts;
|
|
302
|
+
}
|
|
303
|
+
|
|
304
|
+
function resolveGemma4ExpertIntermediateSize(ctx) {
|
|
305
|
+
const manifest = ctx.manifest ?? {};
|
|
306
|
+
const expertIntermediateSize = manifest.moeConfig?.expertIntermediateSize ?? null;
|
|
307
|
+
if (expertIntermediateSize == null) {
|
|
308
|
+
const modelId = manifest.modelId ?? 'unknown';
|
|
309
|
+
throw new Error(`[MoE] Gemma-style manifest "${modelId}" missing moeConfig.expertIntermediateSize`);
|
|
310
|
+
}
|
|
311
|
+
return expertIntermediateSize;
|
|
312
|
+
}
|
|
313
|
+
|
|
314
|
+
async function loadFirstTensor(ctx, names) {
|
|
315
|
+
for (const name of names) {
|
|
316
|
+
const tensor = await ctx.loadTensor(name);
|
|
317
|
+
if (tensor) return tensor;
|
|
318
|
+
}
|
|
319
|
+
return null;
|
|
320
|
+
}
|
|
321
|
+
|
|
190
322
|
function assertMixtralWeights(weights, layerIdx, expertIdx) {
|
|
191
323
|
const missing = [];
|
|
192
324
|
if (!weights.gate) missing.push('gate');
|
|
@@ -213,6 +345,43 @@ function assertGptOssWeights(weights, layerIdx, expertIdx) {
|
|
|
213
345
|
}
|
|
214
346
|
}
|
|
215
347
|
|
|
348
|
+
function assertGemma4Weights(weights, layerIdx, expertIdx) {
|
|
349
|
+
const missing = [];
|
|
350
|
+
if (!weights.gateUp) missing.push('experts.gate_up_proj');
|
|
351
|
+
if (!weights.down) missing.push('experts.down_proj');
|
|
352
|
+
if (!weights.expertIntermediateSize) missing.push('expertIntermediateSize');
|
|
353
|
+
if (missing.length > 0) {
|
|
354
|
+
throw new Error(
|
|
355
|
+
`[MoE] Gemma-style expert ${layerIdx}_${expertIdx} missing tensors: ${missing.join(', ')}`
|
|
356
|
+
);
|
|
357
|
+
}
|
|
358
|
+
}
|
|
359
|
+
|
|
360
|
+
async function loadGemma4StyleExpert(ctx, layerIdx, expertIdx) {
|
|
361
|
+
const packedKey = `layer_${layerIdx}_gemma4_packed`;
|
|
362
|
+
let packed = ctx.experts.get(packedKey);
|
|
363
|
+
|
|
364
|
+
if (!packed) {
|
|
365
|
+
packed = {
|
|
366
|
+
expertFormat: 'gemma4',
|
|
367
|
+
numExperts: resolveGemma4NumExperts(ctx),
|
|
368
|
+
expertIntermediateSize: resolveGemma4ExpertIntermediateSize(ctx),
|
|
369
|
+
gateUp: await loadFirstTensor(ctx, getGemma4ExpertTensorCandidates(layerIdx)[0]),
|
|
370
|
+
down: await loadFirstTensor(ctx, getGemma4ExpertTensorCandidates(layerIdx)[1]),
|
|
371
|
+
};
|
|
372
|
+
ctx.experts.set(packedKey, packed);
|
|
373
|
+
}
|
|
374
|
+
|
|
375
|
+
return {
|
|
376
|
+
expertFormat: 'gemma4',
|
|
377
|
+
expertIdx,
|
|
378
|
+
numExperts: packed.numExperts,
|
|
379
|
+
expertIntermediateSize: packed.expertIntermediateSize,
|
|
380
|
+
gateUp: packed.gateUp,
|
|
381
|
+
down: packed.down,
|
|
382
|
+
};
|
|
383
|
+
}
|
|
384
|
+
|
|
216
385
|
|
|
217
386
|
async function loadGptOssStyleExpert(ctx, layerIdx, expertIdx) {
|
|
218
387
|
const gptOssPrefix = `model.layers.${layerIdx}.mlp.experts`;
|
|
@@ -19,6 +19,7 @@ import type {
|
|
|
19
19
|
ManifestEmbeddingProjectionSchema,
|
|
20
20
|
} from '../config/schema/index.js';
|
|
21
21
|
import type { TensorLocation } from './loader-types.js';
|
|
22
|
+
import type { DiffusionGemmaSelfConditioningWeights } from './loader-types.js';
|
|
22
23
|
|
|
23
24
|
/** Tensor loading function signature */
|
|
24
25
|
export type TensorLoader = (
|
|
@@ -55,6 +56,8 @@ export interface FinalWeightsContext {
|
|
|
55
56
|
embeddings: GPUBuffer | WeightBuffer | CpuWeightBuffer | Float32Array | null;
|
|
56
57
|
/** Optional embedding postprocessor declared by manifest */
|
|
57
58
|
embeddingPostprocessor?: ManifestEmbeddingPostprocessorSchema | null;
|
|
59
|
+
/** Manifest-owned DiffusionGemma self-conditioning enablement */
|
|
60
|
+
diffusionGemmaSelfConditioning?: boolean;
|
|
58
61
|
/** Manifest model type */
|
|
59
62
|
modelType?: string | null;
|
|
60
63
|
/** Whether LM head should fall back to tied embeddings */
|
|
@@ -75,6 +78,8 @@ export interface FinalWeightsResult {
|
|
|
75
78
|
lmHead: GPUBuffer | WeightBuffer | CpuWeightBuffer | Float32Array | null;
|
|
76
79
|
/** Optional embedding-only postprocessor weights */
|
|
77
80
|
embeddingPostprocessor: LoadedEmbeddingPostprocessor | null;
|
|
81
|
+
/** Optional DiffusionGemma decoder self-conditioning weights */
|
|
82
|
+
diffusionGemmaSelfConditioning: DiffusionGemmaSelfConditioningWeights | null;
|
|
78
83
|
/** Whether norm offset debug was logged */
|
|
79
84
|
normOffsetDebugLogged: boolean;
|
|
80
85
|
}
|
|
@@ -5,6 +5,8 @@ import {
|
|
|
5
5
|
createCpuWeightBuffer,
|
|
6
6
|
isWeightBuffer,
|
|
7
7
|
isCpuWeightBuffer,
|
|
8
|
+
isSplitWeightBuffer,
|
|
9
|
+
getWeightDtype,
|
|
8
10
|
} from '../gpu/weight-buffer.js';
|
|
9
11
|
import { maybeDowncastToF16 } from './weight-downcast.js';
|
|
10
12
|
import { getTensorNamesByRole } from './tensors/tensor-role.js';
|
|
@@ -21,6 +23,12 @@ const HEAD_GROUP = 'head';
|
|
|
21
23
|
const FINAL_NORM_ROLE = 'norm';
|
|
22
24
|
const LM_HEAD_ROLE = 'lm_head';
|
|
23
25
|
const EMBEDDING_MODEL_TYPE = 'embedding';
|
|
26
|
+
const DENSE_TIED_LM_HEAD_DTYPES = new Set(['F16', 'F32']);
|
|
27
|
+
const DIFFUSION_GEMMA_SELF_CONDITIONING_PREFIXES = [
|
|
28
|
+
'model.decoder.self_conditioning',
|
|
29
|
+
'decoder.self_conditioning',
|
|
30
|
+
'self_conditioning',
|
|
31
|
+
];
|
|
24
32
|
|
|
25
33
|
function isGpuBufferInstance(value) {
|
|
26
34
|
return typeof GPUBuffer !== 'undefined' && value instanceof GPUBuffer;
|
|
@@ -54,6 +62,77 @@ function createRangeBackedWeightBuffer(ctx, name, location) {
|
|
|
54
62
|
return createCpuWeightBuffer(source, dtype, layout, location.shape, name);
|
|
55
63
|
}
|
|
56
64
|
|
|
65
|
+
function normalizeLocationDtype(location) {
|
|
66
|
+
return typeof location?.dtype === 'string'
|
|
67
|
+
? location.dtype.trim().toUpperCase()
|
|
68
|
+
: '';
|
|
69
|
+
}
|
|
70
|
+
|
|
71
|
+
function isDenseTiedLmHeadLocation(location) {
|
|
72
|
+
if (!location || !DENSE_TIED_LM_HEAD_DTYPES.has(normalizeLocationDtype(location))) {
|
|
73
|
+
return false;
|
|
74
|
+
}
|
|
75
|
+
if (location.sourceTransform || location.storage) {
|
|
76
|
+
return false;
|
|
77
|
+
}
|
|
78
|
+
return true;
|
|
79
|
+
}
|
|
80
|
+
|
|
81
|
+
function getLoadedWeightShape(weight) {
|
|
82
|
+
if (isWeightBuffer(weight) || isCpuWeightBuffer(weight) || isSplitWeightBuffer(weight)) {
|
|
83
|
+
return weight.shape;
|
|
84
|
+
}
|
|
85
|
+
return null;
|
|
86
|
+
}
|
|
87
|
+
|
|
88
|
+
function getLoadedWeightLayout(weight) {
|
|
89
|
+
if (isWeightBuffer(weight) || isCpuWeightBuffer(weight) || isSplitWeightBuffer(weight)) {
|
|
90
|
+
return weight.layout;
|
|
91
|
+
}
|
|
92
|
+
return null;
|
|
93
|
+
}
|
|
94
|
+
|
|
95
|
+
function shapesEqual(left, right) {
|
|
96
|
+
if (!Array.isArray(left) || !Array.isArray(right) || left.length !== right.length) {
|
|
97
|
+
return false;
|
|
98
|
+
}
|
|
99
|
+
for (let i = 0; i < left.length; i++) {
|
|
100
|
+
if (left[i] !== right[i]) {
|
|
101
|
+
return false;
|
|
102
|
+
}
|
|
103
|
+
}
|
|
104
|
+
return true;
|
|
105
|
+
}
|
|
106
|
+
|
|
107
|
+
function shouldUseTiedEmbeddingsForLmHead(ctx, loc) {
|
|
108
|
+
if (ctx.tieWordEmbeddings !== true || !ctx.embeddings) {
|
|
109
|
+
return false;
|
|
110
|
+
}
|
|
111
|
+
if (ctx.embeddings instanceof Float32Array || isCpuWeightBuffer(ctx.embeddings)) {
|
|
112
|
+
return false;
|
|
113
|
+
}
|
|
114
|
+
if (!isDenseTiedLmHeadLocation(loc)) {
|
|
115
|
+
return false;
|
|
116
|
+
}
|
|
117
|
+
|
|
118
|
+
const embeddingShape = getLoadedWeightShape(ctx.embeddings);
|
|
119
|
+
if (!shapesEqual(loc.shape, embeddingShape)) {
|
|
120
|
+
return false;
|
|
121
|
+
}
|
|
122
|
+
|
|
123
|
+
const lmHeadDtype = selectRuleValue('loader', 'weights', 'floatLocationDtype', {
|
|
124
|
+
locationDtype: loc.dtype,
|
|
125
|
+
});
|
|
126
|
+
const embeddingDtype = getWeightDtype(ctx.embeddings);
|
|
127
|
+
if (embeddingDtype !== lmHeadDtype) {
|
|
128
|
+
return false;
|
|
129
|
+
}
|
|
130
|
+
|
|
131
|
+
const lmHeadLayout = ctx.resolveWeightLayout(loc);
|
|
132
|
+
const embeddingLayout = getLoadedWeightLayout(ctx.embeddings);
|
|
133
|
+
return embeddingLayout === lmHeadLayout;
|
|
134
|
+
}
|
|
135
|
+
|
|
57
136
|
function isLikelyFinalNormName(name) {
|
|
58
137
|
const lower = String(name || '').toLowerCase();
|
|
59
138
|
if (!lower) return false;
|
|
@@ -65,6 +144,8 @@ function isLikelyFinalNormName(name) {
|
|
|
65
144
|
|
|
66
145
|
return (
|
|
67
146
|
lower === 'norm.weight' ||
|
|
147
|
+
lower === 'model.decoder.norm.weight' ||
|
|
148
|
+
lower === 'model.encoder.language_model.norm.weight' ||
|
|
68
149
|
lower.includes('model.norm.weight') ||
|
|
69
150
|
lower.includes('language_model.norm.weight') ||
|
|
70
151
|
lower.includes('model.language_model.norm.weight') ||
|
|
@@ -76,6 +157,71 @@ function isLikelyFinalNormName(name) {
|
|
|
76
157
|
);
|
|
77
158
|
}
|
|
78
159
|
|
|
160
|
+
async function loadFirstExistingTensor(ctx, prefixes, suffixes, options = {}) {
|
|
161
|
+
for (const prefix of prefixes) {
|
|
162
|
+
for (const suffix of suffixes) {
|
|
163
|
+
const tensor = await ctx.loadTensor(`${prefix}.${suffix}`, options.toGPU !== false, true);
|
|
164
|
+
if (tensor) return tensor;
|
|
165
|
+
}
|
|
166
|
+
}
|
|
167
|
+
return null;
|
|
168
|
+
}
|
|
169
|
+
|
|
170
|
+
async function loadDiffusionGemmaSelfConditioning(ctx) {
|
|
171
|
+
const preNorm = await loadFirstExistingTensor(
|
|
172
|
+
ctx,
|
|
173
|
+
DIFFUSION_GEMMA_SELF_CONDITIONING_PREFIXES,
|
|
174
|
+
['pre_norm.weight']
|
|
175
|
+
);
|
|
176
|
+
const gateProj = await loadFirstExistingTensor(
|
|
177
|
+
ctx,
|
|
178
|
+
DIFFUSION_GEMMA_SELF_CONDITIONING_PREFIXES,
|
|
179
|
+
['gate_proj.weight']
|
|
180
|
+
);
|
|
181
|
+
const upProj = await loadFirstExistingTensor(
|
|
182
|
+
ctx,
|
|
183
|
+
DIFFUSION_GEMMA_SELF_CONDITIONING_PREFIXES,
|
|
184
|
+
['up_proj.weight']
|
|
185
|
+
);
|
|
186
|
+
const downProj = await loadFirstExistingTensor(
|
|
187
|
+
ctx,
|
|
188
|
+
DIFFUSION_GEMMA_SELF_CONDITIONING_PREFIXES,
|
|
189
|
+
['down_proj.weight']
|
|
190
|
+
);
|
|
191
|
+
const postNorm = await loadFirstExistingTensor(
|
|
192
|
+
ctx,
|
|
193
|
+
DIFFUSION_GEMMA_SELF_CONDITIONING_PREFIXES,
|
|
194
|
+
['post_norm.weight']
|
|
195
|
+
);
|
|
196
|
+
|
|
197
|
+
const anyPresent = preNorm || gateProj || upProj || downProj || postNorm;
|
|
198
|
+
if (!anyPresent) {
|
|
199
|
+
return null;
|
|
200
|
+
}
|
|
201
|
+
const missing = [];
|
|
202
|
+
if (!preNorm) missing.push('pre_norm.weight');
|
|
203
|
+
if (!gateProj) missing.push('gate_proj.weight');
|
|
204
|
+
if (!upProj) missing.push('up_proj.weight');
|
|
205
|
+
if (!downProj) missing.push('down_proj.weight');
|
|
206
|
+
if (missing.length > 0) {
|
|
207
|
+
throw new Error(
|
|
208
|
+
`[Loader] DiffusionGemma self-conditioning weights are incomplete. Missing: ${missing.join(', ')}.`
|
|
209
|
+
);
|
|
210
|
+
}
|
|
211
|
+
|
|
212
|
+
return {
|
|
213
|
+
preNorm,
|
|
214
|
+
postNorm,
|
|
215
|
+
gateProj,
|
|
216
|
+
upProj,
|
|
217
|
+
downProj,
|
|
218
|
+
};
|
|
219
|
+
}
|
|
220
|
+
|
|
221
|
+
function shouldLoadDiffusionGemmaSelfConditioning(ctx) {
|
|
222
|
+
return ctx.diffusionGemmaSelfConditioning === true;
|
|
223
|
+
}
|
|
224
|
+
|
|
79
225
|
// ============================================================================
|
|
80
226
|
// Main Function
|
|
81
227
|
// ============================================================================
|
|
@@ -93,11 +239,15 @@ export async function loadFinalWeights(ctx) {
|
|
|
93
239
|
// Load LM head
|
|
94
240
|
const lmHead = await loadLmHead(ctx);
|
|
95
241
|
const embeddingPostprocessor = await loadEmbeddingPostprocessor(ctx);
|
|
242
|
+
const diffusionGemmaSelfConditioning = shouldLoadDiffusionGemmaSelfConditioning(ctx)
|
|
243
|
+
? await loadDiffusionGemmaSelfConditioning(ctx)
|
|
244
|
+
: null;
|
|
96
245
|
|
|
97
246
|
return {
|
|
98
247
|
finalNorm,
|
|
99
248
|
lmHead,
|
|
100
249
|
embeddingPostprocessor,
|
|
250
|
+
diffusionGemmaSelfConditioning,
|
|
101
251
|
normOffsetDebugLogged,
|
|
102
252
|
};
|
|
103
253
|
}
|
|
@@ -185,6 +335,14 @@ async function loadLmHead(ctx) {
|
|
|
185
335
|
const loc = ctx.tensorLocations.get(name);
|
|
186
336
|
if (!loc) continue;
|
|
187
337
|
|
|
338
|
+
if (shouldUseTiedEmbeddingsForLmHead(ctx, loc)) {
|
|
339
|
+
debugTrace.loader(`Using tied embeddings as dense LM head "${name}" (manifest.tieWordEmbeddings=true)`);
|
|
340
|
+
lmHeadName = name;
|
|
341
|
+
lmHeadLoc = loc;
|
|
342
|
+
lmHead = ctx.embeddings;
|
|
343
|
+
break;
|
|
344
|
+
}
|
|
345
|
+
|
|
188
346
|
const shouldStream = ctx.shouldStreamLargeWeight(name, loc, 'LM head');
|
|
189
347
|
const tensor = shouldStream
|
|
190
348
|
? (
|
|
@@ -192,17 +350,18 @@ async function loadLmHead(ctx) {
|
|
|
192
350
|
?? await ctx.loadTensor(name, false, true)
|
|
193
351
|
)
|
|
194
352
|
: await ctx.loadTensor(name, true, true);
|
|
353
|
+
const tensorShouldStream = shouldStream;
|
|
195
354
|
|
|
196
|
-
if (
|
|
355
|
+
if (tensorShouldStream && tensor && !(tensor instanceof Float32Array) && !isCpuWeightBuffer(tensor)) {
|
|
197
356
|
throw new Error(
|
|
198
357
|
`[Loader] LM head "${name}" too large for GPU and cannot be loaded on CPU (dtype=${loc.dtype}).`
|
|
199
358
|
);
|
|
200
359
|
}
|
|
201
360
|
|
|
202
|
-
if (tensor && (isGpuBufferInstance(tensor) || isWeightBuffer(tensor) || isCpuWeightBuffer(tensor) || tensor instanceof Float32Array)) {
|
|
361
|
+
if (tensor && (isGpuBufferInstance(tensor) || isWeightBuffer(tensor) || isCpuWeightBuffer(tensor) || isSplitWeightBuffer(tensor) || tensor instanceof Float32Array)) {
|
|
203
362
|
lmHeadName = name;
|
|
204
363
|
lmHeadLoc = loc;
|
|
205
|
-
lmHead = processLmHeadTensor(ctx, tensor, name, loc,
|
|
364
|
+
lmHead = processLmHeadTensor(ctx, tensor, name, loc, tensorShouldStream);
|
|
206
365
|
break;
|
|
207
366
|
}
|
|
208
367
|
}
|
|
@@ -287,6 +446,10 @@ async function loadCpuFloatTensor(ctx, tensorName, label) {
|
|
|
287
446
|
|
|
288
447
|
|
|
289
448
|
function processLmHeadTensor(ctx, tensor, name, loc, shouldStream) {
|
|
449
|
+
if (isSplitWeightBuffer(tensor)) {
|
|
450
|
+
return tensor;
|
|
451
|
+
}
|
|
452
|
+
|
|
290
453
|
if (isCpuWeightBuffer(tensor)) {
|
|
291
454
|
log.warn('Loader', `LM head stored on CPU via range-backed source (layout=${tensor.layout})`);
|
|
292
455
|
return tensor;
|
|
@@ -332,7 +495,7 @@ async function maybeDowncastLmHead(ctx, lmHead, lmHeadName, lmHeadLoc) {
|
|
|
332
495
|
}
|
|
333
496
|
|
|
334
497
|
// Can't downcast Float32Array or CpuWeightBuffer
|
|
335
|
-
if (lmHead instanceof Float32Array || isCpuWeightBuffer(lmHead)) {
|
|
498
|
+
if (lmHead instanceof Float32Array || isCpuWeightBuffer(lmHead) || isSplitWeightBuffer(lmHead)) {
|
|
336
499
|
return lmHead;
|
|
337
500
|
}
|
|
338
501
|
|
|
@@ -38,6 +38,8 @@ export interface LayerLoaderContext {
|
|
|
38
38
|
isMoE: boolean;
|
|
39
39
|
/** Check if layer is an expert layer */
|
|
40
40
|
isExpertLayer: (layerIdx: number) => boolean;
|
|
41
|
+
/** Load dense FFN weights even when a layer also has MoE experts. */
|
|
42
|
+
loadDenseFfnForMoeLayers?: boolean;
|
|
41
43
|
/** Optional attention dimensions for fused QKV sizing */
|
|
42
44
|
numHeads?: number;
|
|
43
45
|
numKVHeads?: number;
|
|
@@ -22,6 +22,8 @@ import { dequantizeQ4KM, dequantizeQ4KMRowWise } from '../converter/quantizer.js
|
|
|
22
22
|
|
|
23
23
|
|
|
24
24
|
const LAYER_PREFIXES = (layerIdx) => [
|
|
25
|
+
`model.decoder.layers.${layerIdx}`,
|
|
26
|
+
`model.encoder.language_model.layers.${layerIdx}`,
|
|
25
27
|
`model.language_model.layers.${layerIdx}`,
|
|
26
28
|
`language_model.layers.${layerIdx}`,
|
|
27
29
|
`language_model.model.layers.${layerIdx}`,
|
|
@@ -41,7 +43,10 @@ const ATTN_SUFFIXES = {
|
|
|
41
43
|
kNorm: ['self_attn.k_norm.weight', 'self_attn.k_layernorm.weight', 'attn_k_norm.weight'],
|
|
42
44
|
postAttentionNorm: ['post_attention_layernorm.weight', 'post_attention_norm.weight', 'ffn_norm.weight'],
|
|
43
45
|
preFeedforwardNorm: ['pre_feedforward_layernorm.weight'],
|
|
46
|
+
preFeedforwardNorm2: ['pre_feedforward_layernorm_2.weight'],
|
|
44
47
|
postFeedforwardNorm: ['post_feedforward_layernorm.weight', 'post_ffw_norm.weight'],
|
|
48
|
+
postFeedforwardNorm1: ['post_feedforward_layernorm_1.weight'],
|
|
49
|
+
postFeedforwardNorm2: ['post_feedforward_layernorm_2.weight'],
|
|
45
50
|
postPerLayerInputNorm: ['post_per_layer_input_norm.weight'],
|
|
46
51
|
layerScalar: ['layer_scalar'],
|
|
47
52
|
};
|
|
@@ -76,8 +81,10 @@ const FFN_SUFFIXES = {
|
|
|
76
81
|
|
|
77
82
|
|
|
78
83
|
const ROUTER_SUFFIXES = {
|
|
79
|
-
routerWeight: ['mlp.router.weight', 'block_sparse_moe.gate.weight'],
|
|
84
|
+
routerWeight: ['mlp.router.weight', 'block_sparse_moe.gate.weight', 'router.proj.weight'],
|
|
80
85
|
routerBias: ['mlp.router.bias'],
|
|
86
|
+
routerScale: ['router.scale'],
|
|
87
|
+
routerPerExpertScale: ['router.per_expert_scale'],
|
|
81
88
|
};
|
|
82
89
|
|
|
83
90
|
|
|
@@ -345,7 +352,10 @@ export async function loadLayer(ctx, layerIdx) {
|
|
|
345
352
|
kNorm: null,
|
|
346
353
|
postAttentionNorm: null,
|
|
347
354
|
preFeedforwardNorm: null,
|
|
355
|
+
preFeedforwardNorm2: null,
|
|
348
356
|
postFeedforwardNorm: null,
|
|
357
|
+
postFeedforwardNorm1: null,
|
|
358
|
+
postFeedforwardNorm2: null,
|
|
349
359
|
postNorm: null,
|
|
350
360
|
postAttnNorm: null,
|
|
351
361
|
convInProj: null,
|
|
@@ -368,14 +378,14 @@ export async function loadLayer(ctx, layerIdx) {
|
|
|
368
378
|
// Load attention weights in parallel
|
|
369
379
|
await loadAttentionWeights(ctx, weights, layerIdx, tryLoad, tryLoadNorm);
|
|
370
380
|
|
|
371
|
-
// Load FFN weights (unless MoE expert layer)
|
|
372
|
-
if (!ctx.isMoE || !ctx.isExpertLayer(layerIdx)) {
|
|
381
|
+
// Load FFN weights (unless MoE expert layer owns only routed experts).
|
|
382
|
+
if (!ctx.isMoE || !ctx.isExpertLayer(layerIdx) || ctx.loadDenseFfnForMoeLayers === true) {
|
|
373
383
|
await loadFfnWeights(ctx, weights, layerIdx, tryLoad, prefixes);
|
|
374
384
|
}
|
|
375
385
|
|
|
376
386
|
// Load MoE router weights
|
|
377
387
|
if (ctx.isMoE && ctx.isExpertLayer(layerIdx)) {
|
|
378
|
-
await loadRouterWeights(weights, tryLoad);
|
|
388
|
+
await loadRouterWeights(ctx, weights, layerIdx, tryLoad);
|
|
379
389
|
}
|
|
380
390
|
|
|
381
391
|
// Load attention sinks
|
|
@@ -448,7 +458,10 @@ async function loadAttentionWeights(ctx, weights, layerIdx, tryLoad, tryLoadNorm
|
|
|
448
458
|
kNorm,
|
|
449
459
|
postAttentionNorm,
|
|
450
460
|
preFeedforwardNorm,
|
|
461
|
+
preFeedforwardNorm2,
|
|
451
462
|
postFeedforwardNorm,
|
|
463
|
+
postFeedforwardNorm1,
|
|
464
|
+
postFeedforwardNorm2,
|
|
452
465
|
postPerLayerInputNorm,
|
|
453
466
|
layerScalar,
|
|
454
467
|
convInProj,
|
|
@@ -474,7 +487,10 @@ async function loadAttentionWeights(ctx, weights, layerIdx, tryLoad, tryLoadNorm
|
|
|
474
487
|
tryLoadNorm(ATTN_SUFFIXES.kNorm),
|
|
475
488
|
tryLoadNorm(ATTN_SUFFIXES.postAttentionNorm),
|
|
476
489
|
tryLoadNorm(ATTN_SUFFIXES.preFeedforwardNorm),
|
|
490
|
+
tryLoadNorm(ATTN_SUFFIXES.preFeedforwardNorm2),
|
|
477
491
|
tryLoadNorm(ATTN_SUFFIXES.postFeedforwardNorm),
|
|
492
|
+
tryLoadNorm(ATTN_SUFFIXES.postFeedforwardNorm1),
|
|
493
|
+
tryLoadNorm(ATTN_SUFFIXES.postFeedforwardNorm2),
|
|
478
494
|
tryLoadNorm(ATTN_SUFFIXES.postPerLayerInputNorm),
|
|
479
495
|
tryLoadCpu(ATTN_SUFFIXES.layerScalar),
|
|
480
496
|
tryLoad(CONV_SUFFIXES.convInProj),
|
|
@@ -510,7 +526,10 @@ async function loadAttentionWeights(ctx, weights, layerIdx, tryLoad, tryLoadNorm
|
|
|
510
526
|
|
|
511
527
|
weights.postAttentionNorm = postAttentionNorm;
|
|
512
528
|
weights.preFeedforwardNorm = preFeedforwardNorm;
|
|
529
|
+
weights.preFeedforwardNorm2 = preFeedforwardNorm2;
|
|
513
530
|
weights.postFeedforwardNorm = postFeedforwardNorm;
|
|
531
|
+
weights.postFeedforwardNorm1 = postFeedforwardNorm1;
|
|
532
|
+
weights.postFeedforwardNorm2 = postFeedforwardNorm2;
|
|
514
533
|
weights.postPerLayerInputNorm = postPerLayerInputNorm;
|
|
515
534
|
weights.layerScalar = layerScalar;
|
|
516
535
|
weights.postNorm = weights.postAttentionNorm || weights.preFeedforwardNorm;
|
|
@@ -607,14 +626,63 @@ async function loadStablePerLayerInputGate(ctx, layerIdx, prefixes, tryLoad) {
|
|
|
607
626
|
}
|
|
608
627
|
|
|
609
628
|
|
|
610
|
-
|
|
611
|
-
|
|
629
|
+
function getVectorElementCount(shape, label) {
|
|
630
|
+
if (!Array.isArray(shape) || shape.length === 0) {
|
|
631
|
+
throw new Error(`[LayerLoader] ${label} requires a non-empty shape.`);
|
|
632
|
+
}
|
|
633
|
+
return shape.reduce((product, dim) => {
|
|
634
|
+
if (!Number.isInteger(dim) || dim <= 0) {
|
|
635
|
+
throw new Error(`[LayerLoader] ${label} has invalid shape dimension: ${String(dim)}.`);
|
|
636
|
+
}
|
|
637
|
+
return product * dim;
|
|
638
|
+
}, 1);
|
|
639
|
+
}
|
|
640
|
+
|
|
641
|
+
async function materializeRouterPerExpertScale(ctx, weight, layerIdx) {
|
|
642
|
+
if (!weight || !isWeightBuffer(weight)) return weight;
|
|
643
|
+
const dtype = getWeightDtype(weight);
|
|
644
|
+
if (dtype === 'f32') return weight;
|
|
645
|
+
if (dtype !== 'f16') {
|
|
646
|
+
throw new Error(
|
|
647
|
+
`[LayerLoader] Layer ${layerIdx} router per-expert scale must be f16 or f32, got ${String(dtype)}.`
|
|
648
|
+
);
|
|
649
|
+
}
|
|
650
|
+
|
|
651
|
+
const shape = [...weight.shape];
|
|
652
|
+
const numElements = getVectorElementCount(shape, 'router per-expert scale');
|
|
653
|
+
const inputTensor = createTensor(
|
|
654
|
+
weight.buffer,
|
|
655
|
+
'f16',
|
|
656
|
+
[numElements],
|
|
657
|
+
`${weight.label ?? `layer_${layerIdx}_router_per_expert_scale`}_f16`
|
|
658
|
+
);
|
|
659
|
+
const f32Tensor = await castF16ToF32(inputTensor);
|
|
660
|
+
ctx.gpuBuffers.add(f32Tensor.buffer);
|
|
661
|
+
debugTrace.loader(`Layer ${layerIdx} materialized router per-expert scale F16->F32`);
|
|
662
|
+
|
|
663
|
+
return createWeightBuffer(
|
|
664
|
+
f32Tensor.buffer,
|
|
665
|
+
'f32',
|
|
666
|
+
weight.layout,
|
|
667
|
+
shape,
|
|
668
|
+
weight.label ?? `layer_${layerIdx}_router_per_expert_scale`,
|
|
669
|
+
cloneWeightMaterializations(weight, ['f32']),
|
|
670
|
+
weight.metadata ?? null
|
|
671
|
+
);
|
|
672
|
+
}
|
|
673
|
+
|
|
674
|
+
async function loadRouterWeights(ctx, weights, layerIdx, tryLoad) {
|
|
675
|
+
const [routerWeight, routerBias, routerScale, routerPerExpertScale] = await Promise.all([
|
|
612
676
|
tryLoad(ROUTER_SUFFIXES.routerWeight),
|
|
613
677
|
tryLoad(ROUTER_SUFFIXES.routerBias),
|
|
678
|
+
tryLoad(ROUTER_SUFFIXES.routerScale),
|
|
679
|
+
tryLoad(ROUTER_SUFFIXES.routerPerExpertScale),
|
|
614
680
|
]);
|
|
615
681
|
// Router weights follow matmul dtype/layout rules when present
|
|
616
682
|
weights.routerWeight = (routerWeight);
|
|
617
683
|
weights.routerBias = (routerBias);
|
|
684
|
+
weights.routerScale = (routerScale);
|
|
685
|
+
weights.routerPerExpertScale = (await materializeRouterPerExpertScale(ctx, routerPerExpertScale, layerIdx));
|
|
618
686
|
}
|
|
619
687
|
|
|
620
688
|
// ============================================================================
|