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
|
@@ -0,0 +1,116 @@
|
|
|
1
|
+
enable f16;
|
|
2
|
+
|
|
3
|
+
override WORKGROUP_SIZE: u32 = 256u;
|
|
4
|
+
override COLS_PER_WG: u32 = 8u;
|
|
5
|
+
override LANES_PER_COL: u32 = 32u;
|
|
6
|
+
override TILE_M: u32 = 4u;
|
|
7
|
+
override SCALE_DTYPE: u32 = 1u;
|
|
8
|
+
|
|
9
|
+
const MAX_WORKGROUP_SIZE: u32 = 256u;
|
|
10
|
+
const MAX_TILE_M: u32 = 4u;
|
|
11
|
+
|
|
12
|
+
struct Uniforms {
|
|
13
|
+
M: u32,
|
|
14
|
+
N: u32,
|
|
15
|
+
K: u32,
|
|
16
|
+
alpha: f32,
|
|
17
|
+
transpose_b: u32,
|
|
18
|
+
workgroups_x: u32,
|
|
19
|
+
_pad1: u32,
|
|
20
|
+
_pad2: u32,
|
|
21
|
+
}
|
|
22
|
+
|
|
23
|
+
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
24
|
+
@group(0) @binding(1) var<storage, read> A: array<f16>;
|
|
25
|
+
@group(0) @binding(2) var<storage, read> B_words: array<u32>;
|
|
26
|
+
@group(0) @binding(3) var<storage, read> S_words: array<u32>;
|
|
27
|
+
@group(0) @binding(4) var<storage, read_write> C_f16: array<f16>;
|
|
28
|
+
|
|
29
|
+
var<workgroup> partials: array<f32, MAX_WORKGROUP_SIZE * MAX_TILE_M>;
|
|
30
|
+
|
|
31
|
+
fn packed_byte(byte_index: u32) -> u32 {
|
|
32
|
+
let word = B_words[byte_index >> 2u];
|
|
33
|
+
let shift = (byte_index & 3u) * 8u;
|
|
34
|
+
return (word >> shift) & 0xffu;
|
|
35
|
+
}
|
|
36
|
+
|
|
37
|
+
fn packed_scale_half(scale_index: u32) -> u32 {
|
|
38
|
+
let word = S_words[scale_index >> 1u];
|
|
39
|
+
return select(word & 0xffffu, (word >> 16u) & 0xffffu, (scale_index & 1u) == 1u);
|
|
40
|
+
}
|
|
41
|
+
|
|
42
|
+
fn scale_value(scale_index: u32) -> f32 {
|
|
43
|
+
if (SCALE_DTYPE == 2u) {
|
|
44
|
+
return bitcast<f32>(S_words[scale_index]);
|
|
45
|
+
}
|
|
46
|
+
let half = packed_scale_half(scale_index);
|
|
47
|
+
if (SCALE_DTYPE == 1u) {
|
|
48
|
+
return bitcast<f32>(half << 16u);
|
|
49
|
+
}
|
|
50
|
+
let pair = unpack2x16float(S_words[scale_index >> 1u]);
|
|
51
|
+
return select(pair.x, pair.y, (scale_index & 1u) == 1u);
|
|
52
|
+
}
|
|
53
|
+
|
|
54
|
+
fn offset_binary_int4(byte_value: u32, high_nibble: bool) -> f32 {
|
|
55
|
+
let value = select(byte_value & 0x0fu, (byte_value >> 4u) & 0x0fu, high_nibble);
|
|
56
|
+
return f32(value) - 8.0;
|
|
57
|
+
}
|
|
58
|
+
|
|
59
|
+
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
60
|
+
fn main_batched_f16a(
|
|
61
|
+
@builtin(local_invocation_id) lid: vec3<u32>,
|
|
62
|
+
@builtin(workgroup_id) wg_id: vec3<u32>
|
|
63
|
+
) {
|
|
64
|
+
if (TILE_M > MAX_TILE_M) {
|
|
65
|
+
return;
|
|
66
|
+
}
|
|
67
|
+
|
|
68
|
+
let local_id = lid.x;
|
|
69
|
+
let col_in_wg = local_id / LANES_PER_COL;
|
|
70
|
+
let lane_in_col = local_id - col_in_wg * LANES_PER_COL;
|
|
71
|
+
let col = wg_id.x * COLS_PER_WG + col_in_wg;
|
|
72
|
+
let row_base_out = wg_id.y * TILE_M;
|
|
73
|
+
let groups_per_row = (u.K + 31u) >> 5u;
|
|
74
|
+
|
|
75
|
+
var sums: array<f32, MAX_TILE_M>;
|
|
76
|
+
for (var row_in_tile: u32 = 0u; row_in_tile < TILE_M; row_in_tile = row_in_tile + 1u) {
|
|
77
|
+
sums[row_in_tile] = 0.0;
|
|
78
|
+
}
|
|
79
|
+
|
|
80
|
+
if (col < u.N) {
|
|
81
|
+
let weight_base = col * groups_per_row * 16u;
|
|
82
|
+
let scale_base = col * groups_per_row;
|
|
83
|
+
for (var k: u32 = lane_in_col; k < u.K; k = k + LANES_PER_COL) {
|
|
84
|
+
let group = k >> 5u;
|
|
85
|
+
let lane = k & 31u;
|
|
86
|
+
let byte_value = packed_byte(weight_base + group * 16u + (lane >> 1u));
|
|
87
|
+
let weight = offset_binary_int4(byte_value, (lane & 1u) == 1u) * scale_value(scale_base + group);
|
|
88
|
+
for (var row_in_tile: u32 = 0u; row_in_tile < TILE_M; row_in_tile = row_in_tile + 1u) {
|
|
89
|
+
let row = row_base_out + row_in_tile;
|
|
90
|
+
if (row < u.M) {
|
|
91
|
+
sums[row_in_tile] = sums[row_in_tile] + f32(A[row * u.K + k]) * weight;
|
|
92
|
+
}
|
|
93
|
+
}
|
|
94
|
+
}
|
|
95
|
+
}
|
|
96
|
+
|
|
97
|
+
for (var row_in_tile: u32 = 0u; row_in_tile < TILE_M; row_in_tile = row_in_tile + 1u) {
|
|
98
|
+
partials[row_in_tile * WORKGROUP_SIZE + local_id] = sums[row_in_tile];
|
|
99
|
+
}
|
|
100
|
+
workgroupBarrier();
|
|
101
|
+
|
|
102
|
+
if (lane_in_col == 0u && col < u.N) {
|
|
103
|
+
let lane_base = col_in_wg * LANES_PER_COL;
|
|
104
|
+
for (var row_in_tile: u32 = 0u; row_in_tile < TILE_M; row_in_tile = row_in_tile + 1u) {
|
|
105
|
+
let row = row_base_out + row_in_tile;
|
|
106
|
+
if (row < u.M) {
|
|
107
|
+
var final_sum: f32 = 0.0;
|
|
108
|
+
let partial_base = row_in_tile * WORKGROUP_SIZE + lane_base;
|
|
109
|
+
for (var lane: u32 = 0u; lane < LANES_PER_COL; lane = lane + 1u) {
|
|
110
|
+
final_sum = final_sum + partials[partial_base + lane];
|
|
111
|
+
}
|
|
112
|
+
C_f16[row * u.N + col] = f16(final_sum * u.alpha);
|
|
113
|
+
}
|
|
114
|
+
}
|
|
115
|
+
}
|
|
116
|
+
}
|
|
@@ -0,0 +1,116 @@
|
|
|
1
|
+
enable f16;
|
|
2
|
+
|
|
3
|
+
override WORKGROUP_SIZE: u32 = 256u;
|
|
4
|
+
override COLS_PER_WG: u32 = 8u;
|
|
5
|
+
override LANES_PER_COL: u32 = 32u;
|
|
6
|
+
override TILE_M: u32 = 4u;
|
|
7
|
+
override SCALE_DTYPE: u32 = 1u;
|
|
8
|
+
|
|
9
|
+
const MAX_WORKGROUP_SIZE: u32 = 256u;
|
|
10
|
+
const MAX_TILE_M: u32 = 4u;
|
|
11
|
+
|
|
12
|
+
struct Uniforms {
|
|
13
|
+
M: u32,
|
|
14
|
+
N: u32,
|
|
15
|
+
K: u32,
|
|
16
|
+
alpha: f32,
|
|
17
|
+
transpose_b: u32,
|
|
18
|
+
workgroups_x: u32,
|
|
19
|
+
_pad1: u32,
|
|
20
|
+
_pad2: u32,
|
|
21
|
+
}
|
|
22
|
+
|
|
23
|
+
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
24
|
+
@group(0) @binding(1) var<storage, read> A: array<f16>;
|
|
25
|
+
@group(0) @binding(2) var<storage, read> B_words: array<u32>;
|
|
26
|
+
@group(0) @binding(3) var<storage, read> S_words: array<u32>;
|
|
27
|
+
@group(0) @binding(4) var<storage, read_write> C_f32: array<f32>;
|
|
28
|
+
|
|
29
|
+
var<workgroup> partials: array<f32, MAX_WORKGROUP_SIZE * MAX_TILE_M>;
|
|
30
|
+
|
|
31
|
+
fn packed_byte(byte_index: u32) -> u32 {
|
|
32
|
+
let word = B_words[byte_index >> 2u];
|
|
33
|
+
let shift = (byte_index & 3u) * 8u;
|
|
34
|
+
return (word >> shift) & 0xffu;
|
|
35
|
+
}
|
|
36
|
+
|
|
37
|
+
fn packed_scale_half(scale_index: u32) -> u32 {
|
|
38
|
+
let word = S_words[scale_index >> 1u];
|
|
39
|
+
return select(word & 0xffffu, (word >> 16u) & 0xffffu, (scale_index & 1u) == 1u);
|
|
40
|
+
}
|
|
41
|
+
|
|
42
|
+
fn scale_value(scale_index: u32) -> f32 {
|
|
43
|
+
if (SCALE_DTYPE == 2u) {
|
|
44
|
+
return bitcast<f32>(S_words[scale_index]);
|
|
45
|
+
}
|
|
46
|
+
let half = packed_scale_half(scale_index);
|
|
47
|
+
if (SCALE_DTYPE == 1u) {
|
|
48
|
+
return bitcast<f32>(half << 16u);
|
|
49
|
+
}
|
|
50
|
+
let pair = unpack2x16float(S_words[scale_index >> 1u]);
|
|
51
|
+
return select(pair.x, pair.y, (scale_index & 1u) == 1u);
|
|
52
|
+
}
|
|
53
|
+
|
|
54
|
+
fn offset_binary_int4(byte_value: u32, high_nibble: bool) -> f32 {
|
|
55
|
+
let value = select(byte_value & 0x0fu, (byte_value >> 4u) & 0x0fu, high_nibble);
|
|
56
|
+
return f32(value) - 8.0;
|
|
57
|
+
}
|
|
58
|
+
|
|
59
|
+
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
60
|
+
fn main_batched_f16a_f32out(
|
|
61
|
+
@builtin(local_invocation_id) lid: vec3<u32>,
|
|
62
|
+
@builtin(workgroup_id) wg_id: vec3<u32>
|
|
63
|
+
) {
|
|
64
|
+
if (TILE_M > MAX_TILE_M) {
|
|
65
|
+
return;
|
|
66
|
+
}
|
|
67
|
+
|
|
68
|
+
let local_id = lid.x;
|
|
69
|
+
let col_in_wg = local_id / LANES_PER_COL;
|
|
70
|
+
let lane_in_col = local_id - col_in_wg * LANES_PER_COL;
|
|
71
|
+
let col = wg_id.x * COLS_PER_WG + col_in_wg;
|
|
72
|
+
let row_base_out = wg_id.y * TILE_M;
|
|
73
|
+
let groups_per_row = (u.K + 31u) >> 5u;
|
|
74
|
+
|
|
75
|
+
var sums: array<f32, MAX_TILE_M>;
|
|
76
|
+
for (var row_in_tile: u32 = 0u; row_in_tile < TILE_M; row_in_tile = row_in_tile + 1u) {
|
|
77
|
+
sums[row_in_tile] = 0.0;
|
|
78
|
+
}
|
|
79
|
+
|
|
80
|
+
if (col < u.N) {
|
|
81
|
+
let weight_base = col * groups_per_row * 16u;
|
|
82
|
+
let scale_base = col * groups_per_row;
|
|
83
|
+
for (var k: u32 = lane_in_col; k < u.K; k = k + LANES_PER_COL) {
|
|
84
|
+
let group = k >> 5u;
|
|
85
|
+
let lane = k & 31u;
|
|
86
|
+
let byte_value = packed_byte(weight_base + group * 16u + (lane >> 1u));
|
|
87
|
+
let weight = offset_binary_int4(byte_value, (lane & 1u) == 1u) * scale_value(scale_base + group);
|
|
88
|
+
for (var row_in_tile: u32 = 0u; row_in_tile < TILE_M; row_in_tile = row_in_tile + 1u) {
|
|
89
|
+
let row = row_base_out + row_in_tile;
|
|
90
|
+
if (row < u.M) {
|
|
91
|
+
sums[row_in_tile] = sums[row_in_tile] + f32(A[row * u.K + k]) * weight;
|
|
92
|
+
}
|
|
93
|
+
}
|
|
94
|
+
}
|
|
95
|
+
}
|
|
96
|
+
|
|
97
|
+
for (var row_in_tile: u32 = 0u; row_in_tile < TILE_M; row_in_tile = row_in_tile + 1u) {
|
|
98
|
+
partials[row_in_tile * WORKGROUP_SIZE + local_id] = sums[row_in_tile];
|
|
99
|
+
}
|
|
100
|
+
workgroupBarrier();
|
|
101
|
+
|
|
102
|
+
if (lane_in_col == 0u && col < u.N) {
|
|
103
|
+
let lane_base = col_in_wg * LANES_PER_COL;
|
|
104
|
+
for (var row_in_tile: u32 = 0u; row_in_tile < TILE_M; row_in_tile = row_in_tile + 1u) {
|
|
105
|
+
let row = row_base_out + row_in_tile;
|
|
106
|
+
if (row < u.M) {
|
|
107
|
+
var final_sum: f32 = 0.0;
|
|
108
|
+
let partial_base = row_in_tile * WORKGROUP_SIZE + lane_base;
|
|
109
|
+
for (var lane: u32 = 0u; lane < LANES_PER_COL; lane = lane + 1u) {
|
|
110
|
+
final_sum = final_sum + partials[partial_base + lane];
|
|
111
|
+
}
|
|
112
|
+
C_f32[row * u.N + col] = final_sum * u.alpha;
|
|
113
|
+
}
|
|
114
|
+
}
|
|
115
|
+
}
|
|
116
|
+
}
|
|
@@ -0,0 +1,95 @@
|
|
|
1
|
+
enable f16;
|
|
2
|
+
|
|
3
|
+
override WORKGROUP_SIZE: u32 = 256u;
|
|
4
|
+
override COLS_PER_WG: u32 = 8u;
|
|
5
|
+
override MULTICOL_COLS_PER_WG: u32 = 8u;
|
|
6
|
+
override MULTICOL_THREADS_PER_COL: u32 = 32u;
|
|
7
|
+
override SCALE_DTYPE: u32 = 1u;
|
|
8
|
+
|
|
9
|
+
const MAX_WORKGROUP_SIZE: u32 = 256u;
|
|
10
|
+
|
|
11
|
+
struct Uniforms {
|
|
12
|
+
M: u32,
|
|
13
|
+
N: u32,
|
|
14
|
+
K: u32,
|
|
15
|
+
alpha: f32,
|
|
16
|
+
transpose_b: u32,
|
|
17
|
+
workgroups_x: u32,
|
|
18
|
+
_pad1: u32,
|
|
19
|
+
_pad2: u32,
|
|
20
|
+
}
|
|
21
|
+
|
|
22
|
+
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
23
|
+
@group(0) @binding(1) var<storage, read> A: array<f16>;
|
|
24
|
+
@group(0) @binding(2) var<storage, read> B_words: array<u32>;
|
|
25
|
+
@group(0) @binding(3) var<storage, read> S_words: array<u32>;
|
|
26
|
+
@group(0) @binding(4) var<storage, read_write> C_f16: array<f16>;
|
|
27
|
+
|
|
28
|
+
var<workgroup> partials: array<f32, MAX_WORKGROUP_SIZE>;
|
|
29
|
+
|
|
30
|
+
fn packed_byte(byte_index: u32) -> u32 {
|
|
31
|
+
let word = B_words[byte_index >> 2u];
|
|
32
|
+
let shift = (byte_index & 3u) * 8u;
|
|
33
|
+
return (word >> shift) & 0xffu;
|
|
34
|
+
}
|
|
35
|
+
|
|
36
|
+
fn packed_scale_half(scale_index: u32) -> u32 {
|
|
37
|
+
let word = S_words[scale_index >> 1u];
|
|
38
|
+
return select(word & 0xffffu, (word >> 16u) & 0xffffu, (scale_index & 1u) == 1u);
|
|
39
|
+
}
|
|
40
|
+
|
|
41
|
+
fn scale_value(scale_index: u32) -> f32 {
|
|
42
|
+
if (SCALE_DTYPE == 2u) {
|
|
43
|
+
return bitcast<f32>(S_words[scale_index]);
|
|
44
|
+
}
|
|
45
|
+
let half = packed_scale_half(scale_index);
|
|
46
|
+
if (SCALE_DTYPE == 1u) {
|
|
47
|
+
return bitcast<f32>(half << 16u);
|
|
48
|
+
}
|
|
49
|
+
let pair = unpack2x16float(S_words[scale_index >> 1u]);
|
|
50
|
+
return select(pair.x, pair.y, (scale_index & 1u) == 1u);
|
|
51
|
+
}
|
|
52
|
+
|
|
53
|
+
fn offset_binary_int4(byte_value: u32, high_nibble: bool) -> f32 {
|
|
54
|
+
let value = select(byte_value & 0x0fu, (byte_value >> 4u) & 0x0fu, high_nibble);
|
|
55
|
+
return f32(value) - 8.0;
|
|
56
|
+
}
|
|
57
|
+
|
|
58
|
+
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
59
|
+
fn main_multicol_f16a(
|
|
60
|
+
@builtin(local_invocation_id) lid: vec3<u32>,
|
|
61
|
+
@builtin(workgroup_id) wg_id: vec3<u32>
|
|
62
|
+
) {
|
|
63
|
+
let local_id = lid.x;
|
|
64
|
+
let lanes_per_col = WORKGROUP_SIZE / MULTICOL_COLS_PER_WG;
|
|
65
|
+
let col_in_wg = local_id / lanes_per_col;
|
|
66
|
+
let lane_in_col = local_id - col_in_wg * lanes_per_col;
|
|
67
|
+
let row = wg_id.y;
|
|
68
|
+
let col = wg_id.x * MULTICOL_COLS_PER_WG + col_in_wg;
|
|
69
|
+
let groups_per_row = (u.K + 31u) >> 5u;
|
|
70
|
+
|
|
71
|
+
var sum: f32 = 0.0;
|
|
72
|
+
if (row < u.M && col < u.N) {
|
|
73
|
+
let row_base = col * groups_per_row * 16u;
|
|
74
|
+
let scale_base = col * groups_per_row;
|
|
75
|
+
for (var k: u32 = lane_in_col; k < u.K; k = k + lanes_per_col) {
|
|
76
|
+
let group = k >> 5u;
|
|
77
|
+
let lane = k & 31u;
|
|
78
|
+
let byte_value = packed_byte(row_base + group * 16u + (lane >> 1u));
|
|
79
|
+
let weight = offset_binary_int4(byte_value, (lane & 1u) == 1u) * scale_value(scale_base + group);
|
|
80
|
+
sum = sum + f32(A[row * u.K + k]) * weight;
|
|
81
|
+
}
|
|
82
|
+
}
|
|
83
|
+
|
|
84
|
+
partials[local_id] = sum;
|
|
85
|
+
workgroupBarrier();
|
|
86
|
+
|
|
87
|
+
if (lane_in_col == 0u && row < u.M && col < u.N) {
|
|
88
|
+
var final_sum: f32 = 0.0;
|
|
89
|
+
let base = col_in_wg * lanes_per_col;
|
|
90
|
+
for (var lane: u32 = 0u; lane < lanes_per_col; lane = lane + 1u) {
|
|
91
|
+
final_sum = final_sum + partials[base + lane];
|
|
92
|
+
}
|
|
93
|
+
C_f16[row * u.N + col] = f16(final_sum * u.alpha);
|
|
94
|
+
}
|
|
95
|
+
}
|
|
@@ -0,0 +1,95 @@
|
|
|
1
|
+
enable f16;
|
|
2
|
+
|
|
3
|
+
override WORKGROUP_SIZE: u32 = 256u;
|
|
4
|
+
override COLS_PER_WG: u32 = 8u;
|
|
5
|
+
override MULTICOL_COLS_PER_WG: u32 = 8u;
|
|
6
|
+
override MULTICOL_THREADS_PER_COL: u32 = 32u;
|
|
7
|
+
override SCALE_DTYPE: u32 = 1u;
|
|
8
|
+
|
|
9
|
+
const MAX_WORKGROUP_SIZE: u32 = 256u;
|
|
10
|
+
|
|
11
|
+
struct Uniforms {
|
|
12
|
+
M: u32,
|
|
13
|
+
N: u32,
|
|
14
|
+
K: u32,
|
|
15
|
+
alpha: f32,
|
|
16
|
+
transpose_b: u32,
|
|
17
|
+
workgroups_x: u32,
|
|
18
|
+
_pad1: u32,
|
|
19
|
+
_pad2: u32,
|
|
20
|
+
}
|
|
21
|
+
|
|
22
|
+
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
23
|
+
@group(0) @binding(1) var<storage, read> A: array<f16>;
|
|
24
|
+
@group(0) @binding(2) var<storage, read> B_words: array<u32>;
|
|
25
|
+
@group(0) @binding(3) var<storage, read> S_words: array<u32>;
|
|
26
|
+
@group(0) @binding(4) var<storage, read_write> C_f32: array<f32>;
|
|
27
|
+
|
|
28
|
+
var<workgroup> partials: array<f32, MAX_WORKGROUP_SIZE>;
|
|
29
|
+
|
|
30
|
+
fn packed_byte(byte_index: u32) -> u32 {
|
|
31
|
+
let word = B_words[byte_index >> 2u];
|
|
32
|
+
let shift = (byte_index & 3u) * 8u;
|
|
33
|
+
return (word >> shift) & 0xffu;
|
|
34
|
+
}
|
|
35
|
+
|
|
36
|
+
fn packed_scale_half(scale_index: u32) -> u32 {
|
|
37
|
+
let word = S_words[scale_index >> 1u];
|
|
38
|
+
return select(word & 0xffffu, (word >> 16u) & 0xffffu, (scale_index & 1u) == 1u);
|
|
39
|
+
}
|
|
40
|
+
|
|
41
|
+
fn scale_value(scale_index: u32) -> f32 {
|
|
42
|
+
if (SCALE_DTYPE == 2u) {
|
|
43
|
+
return bitcast<f32>(S_words[scale_index]);
|
|
44
|
+
}
|
|
45
|
+
let half = packed_scale_half(scale_index);
|
|
46
|
+
if (SCALE_DTYPE == 1u) {
|
|
47
|
+
return bitcast<f32>(half << 16u);
|
|
48
|
+
}
|
|
49
|
+
let pair = unpack2x16float(S_words[scale_index >> 1u]);
|
|
50
|
+
return select(pair.x, pair.y, (scale_index & 1u) == 1u);
|
|
51
|
+
}
|
|
52
|
+
|
|
53
|
+
fn offset_binary_int4(byte_value: u32, high_nibble: bool) -> f32 {
|
|
54
|
+
let value = select(byte_value & 0x0fu, (byte_value >> 4u) & 0x0fu, high_nibble);
|
|
55
|
+
return f32(value) - 8.0;
|
|
56
|
+
}
|
|
57
|
+
|
|
58
|
+
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
59
|
+
fn main_multicol_f16a_f32out(
|
|
60
|
+
@builtin(local_invocation_id) lid: vec3<u32>,
|
|
61
|
+
@builtin(workgroup_id) wg_id: vec3<u32>
|
|
62
|
+
) {
|
|
63
|
+
let local_id = lid.x;
|
|
64
|
+
let lanes_per_col = WORKGROUP_SIZE / MULTICOL_COLS_PER_WG;
|
|
65
|
+
let col_in_wg = local_id / lanes_per_col;
|
|
66
|
+
let lane_in_col = local_id - col_in_wg * lanes_per_col;
|
|
67
|
+
let row = wg_id.y;
|
|
68
|
+
let col = wg_id.x * MULTICOL_COLS_PER_WG + col_in_wg;
|
|
69
|
+
let groups_per_row = (u.K + 31u) >> 5u;
|
|
70
|
+
|
|
71
|
+
var sum: f32 = 0.0;
|
|
72
|
+
if (row < u.M && col < u.N) {
|
|
73
|
+
let row_base = col * groups_per_row * 16u;
|
|
74
|
+
let scale_base = col * groups_per_row;
|
|
75
|
+
for (var k: u32 = lane_in_col; k < u.K; k = k + lanes_per_col) {
|
|
76
|
+
let group = k >> 5u;
|
|
77
|
+
let lane = k & 31u;
|
|
78
|
+
let byte_value = packed_byte(row_base + group * 16u + (lane >> 1u));
|
|
79
|
+
let weight = offset_binary_int4(byte_value, (lane & 1u) == 1u) * scale_value(scale_base + group);
|
|
80
|
+
sum = sum + f32(A[row * u.K + k]) * weight;
|
|
81
|
+
}
|
|
82
|
+
}
|
|
83
|
+
|
|
84
|
+
partials[local_id] = sum;
|
|
85
|
+
workgroupBarrier();
|
|
86
|
+
|
|
87
|
+
if (lane_in_col == 0u && row < u.M && col < u.N) {
|
|
88
|
+
var final_sum: f32 = 0.0;
|
|
89
|
+
let base = col_in_wg * lanes_per_col;
|
|
90
|
+
for (var lane: u32 = 0u; lane < lanes_per_col; lane = lane + 1u) {
|
|
91
|
+
final_sum = final_sum + partials[base + lane];
|
|
92
|
+
}
|
|
93
|
+
C_f32[row * u.N + col] = final_sum * u.alpha;
|
|
94
|
+
}
|
|
95
|
+
}
|
|
@@ -0,0 +1,271 @@
|
|
|
1
|
+
// Fused RMSNorm + Q4_K WideTile Matmul — F32 A, Q4_K B, F32 C
|
|
2
|
+
//
|
|
3
|
+
// Prologue fuses RMSNorm into the WideTile matmul:
|
|
4
|
+
// normed[m, k] = (input[m, k] / rms(input[m, :])) * norm_weight[k]
|
|
5
|
+
// output[m, n] = sum_k(normed[m, k] * dequant(B_q4k)[k, n])
|
|
6
|
+
//
|
|
7
|
+
// Targets the two Gemma-style pre-matmul norms:
|
|
8
|
+
// input_norm → {q,k,v}_proj (each q/k/v call runs norm internally)
|
|
9
|
+
// pre_feedforward_norm → {gate,up}_proj
|
|
10
|
+
//
|
|
11
|
+
// Redundant norm work across q/k/v is negligible (norm: O(K) per row;
|
|
12
|
+
// matmul: O(K*N_per_WG) per row). Each saved standalone rmsnorm dispatch
|
|
13
|
+
// cuts ~0.85 ms of Dawn/Vulkan bubble — the target.
|
|
14
|
+
//
|
|
15
|
+
// Weight offset: when `RMS_NORM_OFFSET = true`, the stored norm weight
|
|
16
|
+
// encodes `(weight - 1.0)` (Gemma family convention). This kernel reads
|
|
17
|
+
// stored_w and applies `1.0 + stored_w` at use-site.
|
|
18
|
+
//
|
|
19
|
+
// Shared memory: tileA[TILE_M][QK_K] f32 = 4 KB per WG block (sliced
|
|
20
|
+
// across K in the existing WideTile loop). The full-row normed tensor
|
|
21
|
+
// is also cached in shared at `row_sq_sum` + `rms` derived per row;
|
|
22
|
+
// we compute sum-of-squares across K chunks in the same outer loop.
|
|
23
|
+
//
|
|
24
|
+
// Two-pass structure: pass 1 accumulates sum_sq across K, computes rms;
|
|
25
|
+
// pass 2 reuses the same tileA buffer to multiply normed*weight and
|
|
26
|
+
// dequant+matmul. Pass 2 reads K in the same block iteration order.
|
|
27
|
+
//
|
|
28
|
+
// Adapted from the iter-8 WideTile design.
|
|
29
|
+
|
|
30
|
+
enable f16;
|
|
31
|
+
|
|
32
|
+
const QK_K: u32 = 256u;
|
|
33
|
+
const SUBBLOCK_SIZE: u32 = 32u;
|
|
34
|
+
const NUM_SUBBLOCK_PAIRS: u32 = 4u;
|
|
35
|
+
const U32S_PER_PAIR: u32 = 8u;
|
|
36
|
+
|
|
37
|
+
override TILE_M: u32 = 4u;
|
|
38
|
+
override TILE_N: u32 = 256u;
|
|
39
|
+
override RMS_NORM_OFFSET: bool = false;
|
|
40
|
+
override WEIGHT_IS_F16: bool = false;
|
|
41
|
+
|
|
42
|
+
const MAX_TILE_M: u32 = 8u;
|
|
43
|
+
const MAX_TILE_N: u32 = 256u;
|
|
44
|
+
|
|
45
|
+
struct Uniforms {
|
|
46
|
+
M: u32,
|
|
47
|
+
N: u32,
|
|
48
|
+
K: u32,
|
|
49
|
+
alpha: f32,
|
|
50
|
+
num_blocks_per_row: u32,
|
|
51
|
+
eps: f32,
|
|
52
|
+
_pad0: u32,
|
|
53
|
+
_pad1: u32,
|
|
54
|
+
}
|
|
55
|
+
|
|
56
|
+
struct Q4KBlock {
|
|
57
|
+
d_dmin: u32,
|
|
58
|
+
scales: array<u32, 3>,
|
|
59
|
+
qs: array<u32, 32>,
|
|
60
|
+
}
|
|
61
|
+
|
|
62
|
+
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
63
|
+
@group(0) @binding(1) var<storage, read> A: array<f32>;
|
|
64
|
+
@group(0) @binding(2) var<storage, read> B_q4k: array<Q4KBlock>;
|
|
65
|
+
@group(0) @binding(3) var<storage, read_write> C: array<f32>;
|
|
66
|
+
@group(0) @binding(4) var<storage, read> norm_weight: array<u32>;
|
|
67
|
+
|
|
68
|
+
var<workgroup> tileA: array<f32, MAX_TILE_M * QK_K>;
|
|
69
|
+
// Per-row reduction buffer for pass 1: [TILE_M][TILE_N] partial sum-of-squares.
|
|
70
|
+
// Each thread writes its per-row partial, then workgroup-parallel tree reduction
|
|
71
|
+
// collapses to shared_sum_sq[m][0]. Derived row_rms_recip is cached for pass 2.
|
|
72
|
+
var<workgroup> shared_sum_sq: array<array<f32, MAX_TILE_N>, MAX_TILE_M>;
|
|
73
|
+
var<workgroup> row_rms_recip: array<f32, MAX_TILE_M>;
|
|
74
|
+
|
|
75
|
+
fn unpack_f16_lo(packed: u32) -> f32 { return unpack2x16float(packed).x; }
|
|
76
|
+
fn unpack_f16_hi(packed: u32) -> f32 { return unpack2x16float(packed).y; }
|
|
77
|
+
|
|
78
|
+
// norm_weight binding is declared as array<u32> so it can carry either packed
|
|
79
|
+
// f32 values (bitcast) or packed f16 pairs (Gemma-family hidden weights).
|
|
80
|
+
// Mirrors the load_weight pattern in rmsnorm.wgsl.
|
|
81
|
+
fn load_norm_weight(idx: u32) -> f32 {
|
|
82
|
+
if (WEIGHT_IS_F16) {
|
|
83
|
+
let packed = norm_weight[idx >> 1u];
|
|
84
|
+
let pair = unpack2x16float(packed);
|
|
85
|
+
return select(pair.x, pair.y, (idx & 1u) == 1u);
|
|
86
|
+
}
|
|
87
|
+
return bitcast<f32>(norm_weight[idx]);
|
|
88
|
+
}
|
|
89
|
+
|
|
90
|
+
fn get_scale_byte(scales: array<u32, 3>, byte_idx: u32) -> u32 {
|
|
91
|
+
let word_idx = byte_idx / 4u;
|
|
92
|
+
let byte_in_word = byte_idx % 4u;
|
|
93
|
+
return (scales[word_idx] >> (byte_in_word * 8u)) & 0xFFu;
|
|
94
|
+
}
|
|
95
|
+
|
|
96
|
+
fn get_scale_min_k4(scales: array<u32, 3>, j: u32) -> vec2<u32> {
|
|
97
|
+
var sc: u32;
|
|
98
|
+
var mn: u32;
|
|
99
|
+
if (j < 4u) {
|
|
100
|
+
sc = get_scale_byte(scales, j) & 63u;
|
|
101
|
+
mn = get_scale_byte(scales, j + 4u) & 63u;
|
|
102
|
+
} else {
|
|
103
|
+
let q_j = get_scale_byte(scales, j + 4u);
|
|
104
|
+
let q_lo = get_scale_byte(scales, j - 4u);
|
|
105
|
+
let q_hi = get_scale_byte(scales, j);
|
|
106
|
+
sc = (q_j & 0xFu) | ((q_lo >> 6u) << 4u);
|
|
107
|
+
mn = (q_j >> 4u) | ((q_hi >> 6u) << 4u);
|
|
108
|
+
}
|
|
109
|
+
return vec2<u32>(sc, mn);
|
|
110
|
+
}
|
|
111
|
+
|
|
112
|
+
fn load_normed_a_vec4(row_local: u32, base_elem: u32) -> vec4<f32> {
|
|
113
|
+
// tileA at this point holds normed values (pass 2). The rms scaling
|
|
114
|
+
// and norm_weight multiplication have already been applied when tileA
|
|
115
|
+
// was populated for this block. See $main pass-2 section.
|
|
116
|
+
let a_row_base = row_local * QK_K;
|
|
117
|
+
return vec4<f32>(
|
|
118
|
+
tileA[a_row_base + base_elem + 0u],
|
|
119
|
+
tileA[a_row_base + base_elem + 1u],
|
|
120
|
+
tileA[a_row_base + base_elem + 2u],
|
|
121
|
+
tileA[a_row_base + base_elem + 3u]
|
|
122
|
+
);
|
|
123
|
+
}
|
|
124
|
+
|
|
125
|
+
@compute @workgroup_size(TILE_N, 1, 1)
|
|
126
|
+
fn main(
|
|
127
|
+
@builtin(local_invocation_id) lid: vec3<u32>,
|
|
128
|
+
@builtin(workgroup_id) wg_id: vec3<u32>
|
|
129
|
+
) {
|
|
130
|
+
if (TILE_M > MAX_TILE_M) { return; }
|
|
131
|
+
|
|
132
|
+
let col_local = lid.x;
|
|
133
|
+
let row_base = wg_id.y * TILE_M;
|
|
134
|
+
let col_base = wg_id.x * TILE_N;
|
|
135
|
+
let col = col_base + col_local;
|
|
136
|
+
|
|
137
|
+
var results: array<f32, MAX_TILE_M>;
|
|
138
|
+
for (var i: u32 = 0u; i < TILE_M; i = i + 1u) {
|
|
139
|
+
results[i] = 0.0;
|
|
140
|
+
}
|
|
141
|
+
|
|
142
|
+
let num_blocks = u.num_blocks_per_row;
|
|
143
|
+
|
|
144
|
+
// ===== PASS 1: workgroup-parallel sum-of-squares per row across all K =====
|
|
145
|
+
// Each of TILE_N threads accumulates a per-thread partial across blocks,
|
|
146
|
+
// then a tree reduction over TILE_N lanes finalises the sum per row. The
|
|
147
|
+
// prior serialised 4-thread reduction was the iter-21 perf wall; this
|
|
148
|
+
// restores full workgroup parallelism (log2(TILE_N)=8 reduction steps).
|
|
149
|
+
var local_sum_sq: array<f32, MAX_TILE_M>;
|
|
150
|
+
for (var m: u32 = 0u; m < TILE_M; m = m + 1u) {
|
|
151
|
+
local_sum_sq[m] = 0.0;
|
|
152
|
+
}
|
|
153
|
+
|
|
154
|
+
// Streaming accumulation: thread col_local handles k_in_block == col_local
|
|
155
|
+
// across every block. QK_K (256) == TILE_N (256), so the mapping is 1:1
|
|
156
|
+
// and every lane picks up a K-value per block without serialisation.
|
|
157
|
+
for (var b: u32 = 0u; b < num_blocks; b = b + 1u) {
|
|
158
|
+
let global_k = b * QK_K + col_local;
|
|
159
|
+
if (global_k < u.K) {
|
|
160
|
+
for (var m: u32 = 0u; m < TILE_M; m = m + 1u) {
|
|
161
|
+
let global_row = row_base + m;
|
|
162
|
+
if (global_row < u.M) {
|
|
163
|
+
let v = A[global_row * u.K + global_k];
|
|
164
|
+
local_sum_sq[m] = local_sum_sq[m] + v * v;
|
|
165
|
+
}
|
|
166
|
+
}
|
|
167
|
+
}
|
|
168
|
+
}
|
|
169
|
+
|
|
170
|
+
// Publish partials into shared memory for tree reduction.
|
|
171
|
+
for (var m: u32 = 0u; m < TILE_M; m = m + 1u) {
|
|
172
|
+
shared_sum_sq[m][col_local] = local_sum_sq[m];
|
|
173
|
+
}
|
|
174
|
+
workgroupBarrier();
|
|
175
|
+
|
|
176
|
+
// Tree reduction over TILE_N lanes. All TILE_M rows are reduced in
|
|
177
|
+
// parallel at each stride — 8 barriers total (log2(256)).
|
|
178
|
+
for (var stride: u32 = TILE_N / 2u; stride > 0u; stride = stride >> 1u) {
|
|
179
|
+
if (col_local < stride) {
|
|
180
|
+
for (var m: u32 = 0u; m < TILE_M; m = m + 1u) {
|
|
181
|
+
shared_sum_sq[m][col_local] =
|
|
182
|
+
shared_sum_sq[m][col_local]
|
|
183
|
+
+ shared_sum_sq[m][col_local + stride];
|
|
184
|
+
}
|
|
185
|
+
}
|
|
186
|
+
workgroupBarrier();
|
|
187
|
+
}
|
|
188
|
+
|
|
189
|
+
// Derive 1/rms per row. One thread per row reads the reduced total.
|
|
190
|
+
if (col_local < TILE_M) {
|
|
191
|
+
let mean_sq = shared_sum_sq[col_local][0] / f32(u.K);
|
|
192
|
+
row_rms_recip[col_local] = 1.0 / sqrt(mean_sq + u.eps);
|
|
193
|
+
}
|
|
194
|
+
workgroupBarrier();
|
|
195
|
+
|
|
196
|
+
// ===== PASS 2: reload A, apply rmsnorm + weight, matmul =====
|
|
197
|
+
for (var b: u32 = 0u; b < num_blocks; b = b + 1u) {
|
|
198
|
+
// Cooperative load of tileA raw (re-read from global — cheap vs
|
|
199
|
+
// matmul work), apply norm in place, then matmul. Applying norm
|
|
200
|
+
// in the same cooperative load avoids an extra barrier.
|
|
201
|
+
let total_a_elems = TILE_M * QK_K;
|
|
202
|
+
for (var idx: u32 = col_local; idx < total_a_elems; idx = idx + TILE_N) {
|
|
203
|
+
let load_row_local = idx / QK_K;
|
|
204
|
+
let load_k = idx % QK_K;
|
|
205
|
+
let global_row = row_base + load_row_local;
|
|
206
|
+
let global_k = b * QK_K + load_k;
|
|
207
|
+
var raw: f32 = 0.0;
|
|
208
|
+
if (global_row < u.M && global_k < u.K) {
|
|
209
|
+
raw = A[global_row * u.K + global_k];
|
|
210
|
+
}
|
|
211
|
+
// rmsnorm: x * (1/rms)
|
|
212
|
+
var normed = raw * row_rms_recip[load_row_local];
|
|
213
|
+
// weight: multiply by norm_weight (with Gemma (1+w) offset).
|
|
214
|
+
if (global_k < u.K) {
|
|
215
|
+
let stored_w = load_norm_weight(global_k);
|
|
216
|
+
let effective_w = select(stored_w, 1.0 + stored_w, RMS_NORM_OFFSET);
|
|
217
|
+
normed = normed * effective_w;
|
|
218
|
+
}
|
|
219
|
+
tileA[idx] = normed;
|
|
220
|
+
}
|
|
221
|
+
workgroupBarrier();
|
|
222
|
+
|
|
223
|
+
if (col < u.N) {
|
|
224
|
+
let block = B_q4k[col * num_blocks + b];
|
|
225
|
+
let d = unpack_f16_lo(block.d_dmin);
|
|
226
|
+
let dmin = unpack_f16_hi(block.d_dmin);
|
|
227
|
+
|
|
228
|
+
for (var p: u32 = 0u; p < NUM_SUBBLOCK_PAIRS; p = p + 1u) {
|
|
229
|
+
let sm_even = get_scale_min_k4(block.scales, 2u * p);
|
|
230
|
+
let sm_odd = get_scale_min_k4(block.scales, 2u * p + 1u);
|
|
231
|
+
let scale_even = d * f32(sm_even.x);
|
|
232
|
+
let min_even_scalar = dmin * f32(sm_even.y);
|
|
233
|
+
let scale_odd = d * f32(sm_odd.x);
|
|
234
|
+
let min_odd_scalar = dmin * f32(sm_odd.y);
|
|
235
|
+
let min_even = vec4<f32>(min_even_scalar);
|
|
236
|
+
let min_odd = vec4<f32>(min_odd_scalar);
|
|
237
|
+
|
|
238
|
+
let pair_base_even = p * 64u;
|
|
239
|
+
let pair_base_odd = p * 64u + SUBBLOCK_SIZE;
|
|
240
|
+
let qs_base = p * U32S_PER_PAIR;
|
|
241
|
+
|
|
242
|
+
for (var uidx: u32 = 0u; uidx < U32S_PER_PAIR; uidx = uidx + 1u) {
|
|
243
|
+
let packed = block.qs[qs_base + uidx];
|
|
244
|
+
let lower_u32 = unpack4xU8(packed & 0x0F0F0F0Fu);
|
|
245
|
+
let upper_u32 = unpack4xU8((packed >> 4u) & 0x0F0F0F0Fu);
|
|
246
|
+
let w_even = scale_even * vec4<f32>(lower_u32) - min_even;
|
|
247
|
+
let w_odd = scale_odd * vec4<f32>(upper_u32) - min_odd;
|
|
248
|
+
|
|
249
|
+
let elem_even = pair_base_even + uidx * 4u;
|
|
250
|
+
let elem_odd = pair_base_odd + uidx * 4u;
|
|
251
|
+
|
|
252
|
+
for (var m: u32 = 0u; m < TILE_M; m = m + 1u) {
|
|
253
|
+
let a_even = load_normed_a_vec4(m, elem_even);
|
|
254
|
+
let a_odd = load_normed_a_vec4(m, elem_odd);
|
|
255
|
+
results[m] = results[m] + dot(a_even, w_even) + dot(a_odd, w_odd);
|
|
256
|
+
}
|
|
257
|
+
}
|
|
258
|
+
}
|
|
259
|
+
}
|
|
260
|
+
workgroupBarrier();
|
|
261
|
+
}
|
|
262
|
+
|
|
263
|
+
if (col < u.N) {
|
|
264
|
+
for (var m: u32 = 0u; m < TILE_M; m = m + 1u) {
|
|
265
|
+
let row = row_base + m;
|
|
266
|
+
if (row < u.M) {
|
|
267
|
+
C[row * u.N + col] = results[m] * u.alpha;
|
|
268
|
+
}
|
|
269
|
+
}
|
|
270
|
+
}
|
|
271
|
+
}
|