doppler-gpu 0.4.3 → 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 +28 -0
- package/README.md +54 -33
- package/assets/doppler.svg +57 -1601
- package/package.json +36 -5
- package/src/cli/cli-output.js +34 -1
- package/src/cli/doppler-cli.d.ts +18 -0
- package/src/cli/doppler-cli.js +68 -35
- package/src/cli/doppler-serve.d.ts +210 -0
- package/src/cli/doppler-serve.js +468 -42
- package/src/client/doppler-registry.json +3 -5
- package/src/client/provider.d.ts +1 -1
- package/src/client/provider.js +1 -1
- package/src/client/runtime/index.js +13 -0
- package/src/client/runtime/lora.d.ts +11 -1
- package/src/client/runtime/lora.js +28 -7
- package/src/client/runtime/model-manager.js +35 -11
- package/src/client/runtime/model-session.d.ts +2 -1
- package/src/client/runtime/model-session.js +2 -2
- package/src/client/runtime/model-source.js +20 -1
- package/src/client/runtime/types.d.ts +8 -0
- package/src/client/wrap-pipeline-handle.d.ts +62 -0
- package/src/client/wrap-pipeline-handle.js +188 -13
- package/src/config/conversion/README.md +5 -0
- 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 +3 -3
- package/src/config/conversion/gemma3/gemma-3-1b-it-f16-af32.json +4 -4
- package/src/config/conversion/gemma3/gemma-3-1b-it-q4k-ehf16-af32.json +4 -4
- 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 +17 -9
- package/src/config/conversion/gemma3/translategemma-4b-1b-enes-q4k-ehf16-af32.json +4 -4
- package/src/config/conversion/gemma3/translategemma-4b-it-q4k-ehf16-af32.json +4 -4
- 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 +5 -5
- package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af32.json +4 -4
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af16-int4ple.json +14 -12
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32-int4ple.json +7 -7
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32.json +7 -7
- package/src/config/conversion/gemma4/gemma-4-moe-q4k-ehf16-af32.json +3 -3
- package/src/config/conversion/gpt-oss-20b-f16-xmxfp4.json +3 -3
- package/src/config/conversion/janus/janus-pro-1b-text-q4k-ehaf16.json +3 -3
- package/src/config/conversion/lfm2/lfm2.5-1.2b-instruct-q4k-ehf16-af32.json +3 -3
- package/src/config/conversion/qwen3/qwen-3-5-0-8b-q4k-ehaf16.json +8 -4
- package/src/config/conversion/qwen3/qwen-3-5-2b-q4k-ehaf16.json +8 -4
- package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-eaf16.json +9 -5
- package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-ehaf16.json +8 -4
- package/src/config/goal-completion-matrix.json +552 -0
- package/src/config/kernel-path-loader.js +192 -38
- package/src/config/kernels/kernel-ref-digests.js +52 -18
- package/src/config/kernels/registry.json +3017 -744
- package/src/config/merge-contract-check.js +7 -3
- package/src/config/merge-helpers.d.ts +2 -0
- package/src/config/merge-helpers.js +1 -0
- package/src/config/merge.d.ts +1 -0
- package/src/config/merge.js +30 -0
- 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-validator.js +4 -0
- package/src/config/runtime/experiments/verify/lfm2-verify.json +3 -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/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 +11 -0
- package/src/config/schema/converter.schema.js +8 -0
- package/src/config/schema/doppler.schema.d.ts +1 -0
- package/src/config/schema/doppler.schema.js +38 -6
- package/src/config/schema/execution-v1.schema.d.ts +27 -0
- package/src/config/schema/execution-v1.schema.js +28 -19
- package/src/config/schema/goal-completion-matrix.schema.json +223 -0
- package/src/config/schema/inference.schema.d.ts +11 -0
- package/src/config/schema/manifest.schema.d.ts +107 -1
- package/src/config/schema/manifest.schema.js +10 -1
- 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 -1
- package/src/config/schema/policy-schema-registry.json +26 -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/source-packages/litert/gemma-4-12b-it.json +390 -0
- package/src/config/source-packages/litert/gemma-4-e2b-it.json +7 -7
- 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.json +5 -0
- package/src/config/transforms/capability-transform-resolver.d.ts +13 -2
- package/src/config/transforms/execution-graph-transforms.d.ts +42 -0
- package/src/config/transforms/execution-graph-transforms.js +227 -230
- package/src/converter/conversion-plan.d.ts +2 -0
- package/src/converter/conversion-plan.js +18 -4
- package/src/converter/core.d.ts +8 -0
- package/src/converter/core.js +574 -12
- package/src/converter/quantization-info.js +179 -2
- package/src/converter/quantizer.d.ts +5 -4
- package/src/converter/quantizer.js +16 -2
- package/src/experimental/adapters/lora-loader.js +109 -24
- 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/training/README.md +11 -0
- package/src/experimental/training/autograd.js +33 -8
- 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/distillation/student-fixture.d.ts +5 -0
- package/src/experimental/training/distillation/student-fixture.js +290 -32
- 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 +1737 -49
- package/src/experimental/training/lora.js +23 -1
- package/src/experimental/training/loss.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-command.js +310 -51
- package/src/experimental/training/runner.d.ts +1 -0
- package/src/experimental/training/runner.js +4 -0
- 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 +186 -6
- package/src/experimental/training/workloads.d.ts +37 -0
- package/src/experimental/training/workloads.js +110 -1
- package/src/formats/rdrr/functional-descriptor.d.ts +73 -0
- package/src/formats/rdrr/functional-descriptor.js +230 -0
- package/src/formats/rdrr/index.d.ts +1 -0
- package/src/formats/rdrr/index.js +1 -0
- package/src/formats/rdrr/parsing.js +22 -6
- package/src/formats/rdrr/source-transform-contract.js +10 -1
- package/src/formats/rdrr/storage-descriptor.js +9 -2
- package/src/formats/rdrr/types.d.ts +5 -3
- package/src/formats/rdrr/validation.js +15 -1
- package/src/gpu/command-recorder.d.ts +43 -3
- package/src/gpu/command-recorder.js +122 -25
- package/src/gpu/device.js +111 -11
- package/src/gpu/kernel-tuner/cache.js +1 -5
- package/src/gpu/kernels/attention.js +17 -9
- package/src/gpu/kernels/attention_decode_online_head256_f16kv.wgsl +236 -0
- package/src/gpu/kernels/attention_streaming_f16.wgsl +7 -5
- 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/cast.js +74 -67
- package/src/gpu/kernels/codegen/patches/attention_streaming_f16.from.attention_streaming.diff +46 -1
- package/src/gpu/kernels/codegen/patches/gelu_f16.from.gelu.diff +0 -20
- package/src/gpu/kernels/codegen/patches/residual_f16.from.residual.diff +15 -12
- package/src/gpu/kernels/codegen/patches/rmsnorm_f16.from.rmsnorm.diff +59 -44
- 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.js +13 -3
- package/src/gpu/kernels/fused_ffn.js +3 -4
- 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_rmsnorm.js +101 -80
- 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/gather.d.ts +84 -1
- package/src/gpu/kernels/gather.js +353 -5
- 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 +11 -7
- 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/index.d.ts +71 -0
- package/src/gpu/kernels/index.js +52 -0
- 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/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/matmul-dispatch.d.ts +11 -0
- package/src/gpu/kernels/matmul-dispatch.js +37 -1
- package/src/gpu/kernels/matmul-selection.d.ts +5 -5
- package/src/gpu/kernels/matmul-selection.js +104 -9
- package/src/gpu/kernels/matmul.d.ts +1 -1
- package/src/gpu/kernels/matmul.js +151 -43
- package/src/gpu/kernels/moe.d.ts +1 -0
- package/src/gpu/kernels/moe.js +46 -18
- package/src/gpu/kernels/pipeline-cache.d.ts +8 -0
- package/src/gpu/kernels/pipeline-cache.js +24 -0
- 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 -0
- package/src/gpu/kernels/residual.js +13 -1
- package/src/gpu/kernels/residual.wgsl +3 -3
- package/src/gpu/kernels/residual_f16.wgsl +5 -3
- 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 +31 -0
- package/src/gpu/kernels/rmsnorm.js +32 -7
- package/src/gpu/kernels/rmsnorm.wgsl +10 -6
- package/src/gpu/kernels/rmsnorm_f16.wgsl +7 -3
- 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_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.js +212 -187
- package/src/gpu/kernels/scatter_add_dynamic_expert_scale.wgsl +47 -0
- 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.js +1 -21
- 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 +5 -80
- 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_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/uniform-utils.js +2 -5
- package/src/gpu/kernels/utils.d.ts +16 -1
- package/src/gpu/kernels/utils.js +42 -13
- package/src/gpu/readback-utils.js +10 -2
- package/src/gpu/uniform-cache.d.ts +10 -1
- package/src/gpu/uniform-cache.js +25 -7
- package/src/gpu/weight-buffer.d.ts +76 -7
- package/src/gpu/weight-buffer.js +85 -1
- package/src/inference/browser-harness-diffusion-energy-suites.js +269 -1
- package/src/inference/browser-harness-suite-helpers.d.ts +9 -0
- package/src/inference/browser-harness-suite-helpers.js +438 -2
- package/src/inference/browser-harness-text-helpers.d.ts +77 -0
- package/src/inference/browser-harness-text-helpers.js +236 -11
- package/src/inference/browser-harness.d.ts +8 -0
- package/src/inference/browser-harness.js +292 -8
- 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 +25 -0
- package/src/inference/kv-cache/base.js +109 -1
- package/src/inference/kv-cache/mixed-geometry.d.ts +10 -0
- package/src/inference/kv-cache/mixed-geometry.js +109 -2
- package/src/inference/kv-cache/sliding-window.d.ts +14 -0
- package/src/inference/kv-cache/sliding-window.js +88 -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/pipelines/audio/gemma4.js +52 -1
- 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/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 +71 -2
- package/src/inference/pipelines/text/attention/projections.js +329 -49
- package/src/inference/pipelines/text/attention/quant-options.d.ts +38 -0
- package/src/inference/pipelines/text/attention/record.js +446 -73
- package/src/inference/pipelines/text/attention/run.js +251 -49
- package/src/inference/pipelines/text/attention/types.d.ts +5 -1
- package/src/inference/pipelines/text/config.d.ts +22 -1
- package/src/inference/pipelines/text/config.js +199 -22
- 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 +4 -3
- package/src/inference/pipelines/text/embed.js +16 -6
- package/src/inference/pipelines/text/execution-plan.d.ts +8 -0
- package/src/inference/pipelines/text/execution-plan.js +6 -0
- package/src/inference/pipelines/text/execution-runtime-builders.d.ts +4 -0
- package/src/inference/pipelines/text/execution-runtime-builders.js +44 -28
- package/src/inference/pipelines/text/execution-v1.d.ts +2 -0
- package/src/inference/pipelines/text/execution-v1.js +373 -139
- package/src/inference/pipelines/text/ffn/dense.js +44 -13
- 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 +44 -4
- package/src/inference/pipelines/text/generator-decode-policy.d.ts +81 -0
- package/src/inference/pipelines/text/generator-decode-policy.js +2 -12
- package/src/inference/pipelines/text/generator-helpers.js +5 -2
- package/src/inference/pipelines/text/generator-prefill-helpers.d.ts +84 -0
- package/src/inference/pipelines/text/generator-prefill-helpers.js +67 -1
- package/src/inference/pipelines/text/generator-runtime.d.ts +1 -0
- package/src/inference/pipelines/text/generator-steps.d.ts +38 -3
- package/src/inference/pipelines/text/generator-steps.js +268 -77
- package/src/inference/pipelines/text/generator.d.ts +39 -0
- package/src/inference/pipelines/text/generator.js +1105 -18
- package/src/inference/pipelines/text/init.d.ts +8 -1
- package/src/inference/pipelines/text/init.js +30 -1
- package/src/inference/pipelines/text/layer-plan-gpu.js +62 -14
- package/src/inference/pipelines/text/layer.js +272 -25
- package/src/inference/pipelines/text/linear-attention.js +6 -4
- package/src/inference/pipelines/text/logits/gpu.d.ts +49 -2
- package/src/inference/pipelines/text/logits/gpu.js +417 -66
- package/src/inference/pipelines/text/logits/index.d.ts +2 -2
- 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/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 +29 -2
- package/src/inference/pipelines/text/ops.js +19 -0
- package/src/inference/pipelines/text/per-layer-inputs.js +5 -0
- package/src/inference/pipelines/text/probes.js +4 -5
- 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.js +13 -9
- package/src/inference/pipelines/text/state.js +17 -0
- package/src/inference/pipelines/text/tsir-fixture-writer.d.ts +57 -0
- package/src/inference/pipelines/text/tsir-fixture-writer.js +8 -9
- package/src/inference/pipelines/text/types.d.ts +84 -3
- package/src/inference/pipelines/text/weights.d.ts +4 -2
- package/src/inference/pipelines/text/weights.js +4 -5
- package/src/inference/pipelines/text.d.ts +24 -0
- package/src/inference/pipelines/text.js +162 -38
- package/src/inference/pipelines/vision/gemma4.js +20 -9
- package/src/inference/tokenizer.d.ts +30 -0
- package/src/inference/tokenizer.js +340 -164
- package/src/inference/tokenizers/bundled.js +71 -49
- package/src/loader/doppler-loader.d.ts +11 -0
- package/src/loader/doppler-loader.js +290 -15
- 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 +263 -8
- 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 +59 -2
- package/src/loader/manifest-config.d.ts +8 -0
- package/src/loader/shard-resolver.js +37 -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/tensor-loader.d.ts +36 -1
- package/src/loader/tensors/tensor-loader.js +437 -21
- 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/buffer-pool.js +12 -9
- package/src/models/diffusiongemma.d.ts +7 -0
- package/src/models/diffusiongemma.js +25 -0
- package/src/models/gemma4.js +24 -0
- package/src/models/qwen3.js +2 -2
- package/src/rules/execution-rules-contract-check.js +93 -10
- package/src/rules/inference/capability-transforms.rules.json +26 -2
- 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 +108 -17
- package/src/rules/inference/ffn.rules.json +18 -0
- 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 +8 -0
- package/src/rules/rule-registry.js +6 -0
- package/src/storage/artifact-storage-context.d.ts +1 -0
- package/src/storage/artifact-storage-context.js +10 -0
- package/src/tooling/browser-command-runner.d.ts +0 -3
- package/src/tooling/browser-command-runner.js +2 -14
- package/src/tooling/command-api-family-normalizers.js +4 -2
- package/src/tooling/command-api-helpers.js +2 -1
- package/src/tooling/command-api.d.ts +2 -0
- package/src/tooling/command-api.js +1 -0
- package/src/tooling/command-runner-shared.d.ts +4 -0
- package/src/tooling/command-runner-shared.js +2 -1
- package/src/tooling/diagnose-runner.js +2 -0
- package/src/tooling/litert-package-runtime.js +553 -16
- package/src/tooling/node-browser-command-runner.js +10 -4
- package/src/tooling/node-command-runner.js +7 -2
- package/src/tooling/node-convert-worker-pool.js +3 -0
- package/src/tooling/node-convert-worker.js +1 -0
- package/src/tooling/node-converter.js +12 -0
- package/src/tooling/node-webgpu.d.ts +1 -0
- package/src/tooling/node-webgpu.js +23 -11
- package/src/tooling/program-bundle.d.ts +52 -1
- package/src/tooling/rdrr-integrity-refresh.d.ts +18 -0
- package/src/tooling/source-artifact-adapter.js +7 -2
- package/src/tooling/source-package-profiles.d.ts +18 -1
- package/src/tooling/source-package-profiles.js +109 -1
- package/src/tooling/source-runtime-bundle.d.ts +1 -0
- package/src/tooling/source-runtime-bundle.js +27 -0
- package/src/tooling/source-runtime-converter-config.js +12 -3
- package/src/tooling-experimental-exports.shared.d.ts +10 -0
- package/src/tooling-experimental-exports.shared.js +10 -0
|
@@ -0,0 +1,146 @@
|
|
|
1
|
+
override WORKGROUP_SIZE: u32 = 256u;
|
|
2
|
+
const MAX_WORKGROUP_SIZE: u32 = 256u;
|
|
3
|
+
|
|
4
|
+
override RMS_NORM_OFFSET: bool = false;
|
|
5
|
+
override Q_WEIGHT_IS_F16: bool = false;
|
|
6
|
+
override K_WEIGHT_IS_F16: bool = false;
|
|
7
|
+
|
|
8
|
+
struct Uniforms {
|
|
9
|
+
num_tokens: u32,
|
|
10
|
+
q_size: u32,
|
|
11
|
+
k_size: u32,
|
|
12
|
+
v_size: u32,
|
|
13
|
+
num_heads: u32,
|
|
14
|
+
num_kv_heads: u32,
|
|
15
|
+
head_dim: u32,
|
|
16
|
+
workgroup_stride: u32,
|
|
17
|
+
qk_rows: u32,
|
|
18
|
+
total_v: u32,
|
|
19
|
+
eps: f32,
|
|
20
|
+
_pad0: u32,
|
|
21
|
+
}
|
|
22
|
+
|
|
23
|
+
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
24
|
+
@group(0) @binding(1) var<storage, read> fused_qkv: array<f32>;
|
|
25
|
+
@group(0) @binding(2) var<storage, read> q_weight: array<u32>;
|
|
26
|
+
@group(0) @binding(3) var<storage, read> k_weight: array<u32>;
|
|
27
|
+
@group(0) @binding(4) var<storage, read_write> q_output: array<f32>;
|
|
28
|
+
@group(0) @binding(5) var<storage, read_write> k_output: array<f32>;
|
|
29
|
+
@group(0) @binding(6) var<storage, read_write> v_output: array<f32>;
|
|
30
|
+
|
|
31
|
+
var<workgroup> shared_sum: array<f32, MAX_WORKGROUP_SIZE>;
|
|
32
|
+
|
|
33
|
+
fn flat_group_index(wg_id: vec3<u32>) -> u32 {
|
|
34
|
+
return wg_id.y * max(u.workgroup_stride, 1u) + wg_id.x;
|
|
35
|
+
}
|
|
36
|
+
|
|
37
|
+
fn q_rows() -> u32 {
|
|
38
|
+
return u.num_tokens * u.num_heads;
|
|
39
|
+
}
|
|
40
|
+
|
|
41
|
+
fn qkv_size() -> u32 {
|
|
42
|
+
return u.q_size + u.k_size + u.v_size;
|
|
43
|
+
}
|
|
44
|
+
|
|
45
|
+
fn apply_weight(w: f32) -> f32 {
|
|
46
|
+
if (RMS_NORM_OFFSET) {
|
|
47
|
+
return 1.0 + w;
|
|
48
|
+
}
|
|
49
|
+
return w;
|
|
50
|
+
}
|
|
51
|
+
|
|
52
|
+
fn load_qk_value(row: u32, idx: u32) -> f32 {
|
|
53
|
+
let q_row_count = q_rows();
|
|
54
|
+
if (row < q_row_count) {
|
|
55
|
+
let token = row / u.num_heads;
|
|
56
|
+
let head = row % u.num_heads;
|
|
57
|
+
return fused_qkv[token * qkv_size() + head * u.head_dim + idx];
|
|
58
|
+
}
|
|
59
|
+
|
|
60
|
+
let k_row = row - q_row_count;
|
|
61
|
+
let token = k_row / u.num_kv_heads;
|
|
62
|
+
let head = k_row % u.num_kv_heads;
|
|
63
|
+
return fused_qkv[token * qkv_size() + u.q_size + head * u.head_dim + idx];
|
|
64
|
+
}
|
|
65
|
+
|
|
66
|
+
fn load_weight(row: u32, idx: u32) -> f32 {
|
|
67
|
+
if (row < q_rows()) {
|
|
68
|
+
if (Q_WEIGHT_IS_F16) {
|
|
69
|
+
let packed = q_weight[idx >> 1u];
|
|
70
|
+
let pair = unpack2x16float(packed);
|
|
71
|
+
return select(pair.x, pair.y, (idx & 1u) == 1u);
|
|
72
|
+
}
|
|
73
|
+
return bitcast<f32>(q_weight[idx]);
|
|
74
|
+
}
|
|
75
|
+
if (K_WEIGHT_IS_F16) {
|
|
76
|
+
let packed = k_weight[idx >> 1u];
|
|
77
|
+
let pair = unpack2x16float(packed);
|
|
78
|
+
return select(pair.x, pair.y, (idx & 1u) == 1u);
|
|
79
|
+
}
|
|
80
|
+
return bitcast<f32>(k_weight[idx]);
|
|
81
|
+
}
|
|
82
|
+
|
|
83
|
+
fn store_qk_value(row: u32, idx: u32, value: f32) {
|
|
84
|
+
if (row < q_rows()) {
|
|
85
|
+
q_output[row * u.head_dim + idx] = value;
|
|
86
|
+
} else {
|
|
87
|
+
let k_row = row - q_rows();
|
|
88
|
+
k_output[k_row * u.head_dim + idx] = value;
|
|
89
|
+
}
|
|
90
|
+
}
|
|
91
|
+
|
|
92
|
+
fn copy_v(v_group: u32, thread_idx: u32) {
|
|
93
|
+
let v_idx = v_group * WORKGROUP_SIZE + thread_idx;
|
|
94
|
+
if (v_idx >= u.total_v) {
|
|
95
|
+
return;
|
|
96
|
+
}
|
|
97
|
+
let token = v_idx / u.v_size;
|
|
98
|
+
let elem = v_idx % u.v_size;
|
|
99
|
+
let src_idx = token * qkv_size() + u.q_size + u.k_size + elem;
|
|
100
|
+
v_output[v_idx] = fused_qkv[src_idx];
|
|
101
|
+
}
|
|
102
|
+
|
|
103
|
+
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
104
|
+
fn main(
|
|
105
|
+
@builtin(local_invocation_id) local_id: vec3<u32>,
|
|
106
|
+
@builtin(workgroup_id) wg_id: vec3<u32>
|
|
107
|
+
) {
|
|
108
|
+
let group_idx = flat_group_index(wg_id);
|
|
109
|
+
let thread_idx = local_id.x;
|
|
110
|
+
|
|
111
|
+
if (group_idx >= u.qk_rows) {
|
|
112
|
+
copy_v(group_idx - u.qk_rows, thread_idx);
|
|
113
|
+
return;
|
|
114
|
+
}
|
|
115
|
+
|
|
116
|
+
let elements_per_thread = (u.head_dim + WORKGROUP_SIZE - 1u) / WORKGROUP_SIZE;
|
|
117
|
+
var local_sum_sq: f32 = 0.0;
|
|
118
|
+
for (var i: u32 = 0u; i < elements_per_thread; i = i + 1u) {
|
|
119
|
+
let idx = thread_idx * elements_per_thread + i;
|
|
120
|
+
if (idx < u.head_dim) {
|
|
121
|
+
let x = load_qk_value(group_idx, idx);
|
|
122
|
+
local_sum_sq = local_sum_sq + x * x;
|
|
123
|
+
}
|
|
124
|
+
}
|
|
125
|
+
|
|
126
|
+
shared_sum[thread_idx] = local_sum_sq;
|
|
127
|
+
workgroupBarrier();
|
|
128
|
+
|
|
129
|
+
for (var stride: u32 = WORKGROUP_SIZE / 2u; stride > 0u; stride = stride >> 1u) {
|
|
130
|
+
if (thread_idx < stride) {
|
|
131
|
+
shared_sum[thread_idx] = shared_sum[thread_idx] + shared_sum[thread_idx + stride];
|
|
132
|
+
}
|
|
133
|
+
workgroupBarrier();
|
|
134
|
+
}
|
|
135
|
+
|
|
136
|
+
let inv_rms = 1.0 / sqrt(shared_sum[0] / f32(u.head_dim) + u.eps);
|
|
137
|
+
workgroupBarrier();
|
|
138
|
+
|
|
139
|
+
for (var i: u32 = 0u; i < elements_per_thread; i = i + 1u) {
|
|
140
|
+
let idx = thread_idx * elements_per_thread + i;
|
|
141
|
+
if (idx < u.head_dim) {
|
|
142
|
+
let result = load_qk_value(group_idx, idx) * inv_rms * apply_weight(load_weight(group_idx, idx));
|
|
143
|
+
store_qk_value(group_idx, idx, result);
|
|
144
|
+
}
|
|
145
|
+
}
|
|
146
|
+
}
|
|
@@ -0,0 +1,185 @@
|
|
|
1
|
+
override WORKGROUP_SIZE: u32 = 256u;
|
|
2
|
+
const MAX_WORKGROUP_SIZE: u32 = 256u;
|
|
3
|
+
|
|
4
|
+
override RMS_NORM_OFFSET: bool = false;
|
|
5
|
+
override Q_WEIGHT_IS_F16: bool = false;
|
|
6
|
+
override K_WEIGHT_IS_F16: bool = false;
|
|
7
|
+
|
|
8
|
+
struct Uniforms {
|
|
9
|
+
num_tokens: u32,
|
|
10
|
+
q_size: u32,
|
|
11
|
+
k_size: u32,
|
|
12
|
+
v_size: u32,
|
|
13
|
+
num_heads: u32,
|
|
14
|
+
num_kv_heads: u32,
|
|
15
|
+
head_dim: u32,
|
|
16
|
+
workgroup_stride: u32,
|
|
17
|
+
qk_rows: u32,
|
|
18
|
+
total_v: u32,
|
|
19
|
+
start_pos: u32,
|
|
20
|
+
half_dim: u32,
|
|
21
|
+
eps: f32,
|
|
22
|
+
_pad0: u32,
|
|
23
|
+
_pad1: u32,
|
|
24
|
+
_pad2: u32,
|
|
25
|
+
}
|
|
26
|
+
|
|
27
|
+
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
28
|
+
@group(0) @binding(1) var<storage, read> fused_qkv: array<f32>;
|
|
29
|
+
@group(0) @binding(2) var<storage, read> q_weight: array<u32>;
|
|
30
|
+
@group(0) @binding(3) var<storage, read> k_weight: array<u32>;
|
|
31
|
+
@group(0) @binding(4) var<storage, read> freqs_cos: array<f32>;
|
|
32
|
+
@group(0) @binding(5) var<storage, read> freqs_sin: array<f32>;
|
|
33
|
+
@group(0) @binding(6) var<storage, read_write> q_output: array<f32>;
|
|
34
|
+
@group(0) @binding(7) var<storage, read_write> k_output: array<f32>;
|
|
35
|
+
@group(0) @binding(8) var<storage, read_write> v_output: array<f32>;
|
|
36
|
+
|
|
37
|
+
var<workgroup> shared_sum: array<f32, MAX_WORKGROUP_SIZE>;
|
|
38
|
+
|
|
39
|
+
fn flat_group_index(wg_id: vec3<u32>) -> u32 {
|
|
40
|
+
return wg_id.y * max(u.workgroup_stride, 1u) + wg_id.x;
|
|
41
|
+
}
|
|
42
|
+
|
|
43
|
+
fn q_rows() -> u32 {
|
|
44
|
+
return u.num_tokens * u.num_heads;
|
|
45
|
+
}
|
|
46
|
+
|
|
47
|
+
fn qkv_size() -> u32 {
|
|
48
|
+
return u.q_size + u.k_size + u.v_size;
|
|
49
|
+
}
|
|
50
|
+
|
|
51
|
+
fn apply_weight(w: f32) -> f32 {
|
|
52
|
+
if (RMS_NORM_OFFSET) {
|
|
53
|
+
return 1.0 + w;
|
|
54
|
+
}
|
|
55
|
+
return w;
|
|
56
|
+
}
|
|
57
|
+
|
|
58
|
+
fn token_for_row(row: u32) -> u32 {
|
|
59
|
+
if (row < q_rows()) {
|
|
60
|
+
return row / u.num_heads;
|
|
61
|
+
}
|
|
62
|
+
let k_row = row - q_rows();
|
|
63
|
+
return k_row / u.num_kv_heads;
|
|
64
|
+
}
|
|
65
|
+
|
|
66
|
+
fn load_qk_value(row: u32, idx: u32) -> f32 {
|
|
67
|
+
if (row < q_rows()) {
|
|
68
|
+
let token = row / u.num_heads;
|
|
69
|
+
let head = row % u.num_heads;
|
|
70
|
+
return fused_qkv[token * qkv_size() + head * u.head_dim + idx];
|
|
71
|
+
}
|
|
72
|
+
|
|
73
|
+
let k_row = row - q_rows();
|
|
74
|
+
let token = k_row / u.num_kv_heads;
|
|
75
|
+
let head = k_row % u.num_kv_heads;
|
|
76
|
+
return fused_qkv[token * qkv_size() + u.q_size + head * u.head_dim + idx];
|
|
77
|
+
}
|
|
78
|
+
|
|
79
|
+
fn load_weight(row: u32, idx: u32) -> f32 {
|
|
80
|
+
if (row < q_rows()) {
|
|
81
|
+
if (Q_WEIGHT_IS_F16) {
|
|
82
|
+
let packed = q_weight[idx >> 1u];
|
|
83
|
+
let pair = unpack2x16float(packed);
|
|
84
|
+
return select(pair.x, pair.y, (idx & 1u) == 1u);
|
|
85
|
+
}
|
|
86
|
+
return bitcast<f32>(q_weight[idx]);
|
|
87
|
+
}
|
|
88
|
+
if (K_WEIGHT_IS_F16) {
|
|
89
|
+
let packed = k_weight[idx >> 1u];
|
|
90
|
+
let pair = unpack2x16float(packed);
|
|
91
|
+
return select(pair.x, pair.y, (idx & 1u) == 1u);
|
|
92
|
+
}
|
|
93
|
+
return bitcast<f32>(k_weight[idx]);
|
|
94
|
+
}
|
|
95
|
+
|
|
96
|
+
fn store_qk_value(row: u32, idx: u32, value: f32) {
|
|
97
|
+
if (row < q_rows()) {
|
|
98
|
+
q_output[row * u.head_dim + idx] = value;
|
|
99
|
+
} else {
|
|
100
|
+
let k_row = row - q_rows();
|
|
101
|
+
k_output[k_row * u.head_dim + idx] = value;
|
|
102
|
+
}
|
|
103
|
+
}
|
|
104
|
+
|
|
105
|
+
fn normalized_value(row: u32, idx: u32, inv_rms: f32) -> f32 {
|
|
106
|
+
return load_qk_value(row, idx) * inv_rms * apply_weight(load_weight(row, idx));
|
|
107
|
+
}
|
|
108
|
+
|
|
109
|
+
fn rotate_pair(x0: f32, x1: f32, cos_val: f32, sin_val: f32) -> vec2<f32> {
|
|
110
|
+
return vec2<f32>(
|
|
111
|
+
x0 * cos_val - x1 * sin_val,
|
|
112
|
+
x0 * sin_val + x1 * cos_val
|
|
113
|
+
);
|
|
114
|
+
}
|
|
115
|
+
|
|
116
|
+
fn store_rotated_qk_pair(row: u32, pair_idx: u32, inv_rms: f32) {
|
|
117
|
+
let first_idx = pair_idx;
|
|
118
|
+
let second_idx = pair_idx + u.half_dim;
|
|
119
|
+
let token = token_for_row(row);
|
|
120
|
+
let freq_idx = (u.start_pos + token) * u.half_dim + pair_idx;
|
|
121
|
+
let rotated = rotate_pair(
|
|
122
|
+
normalized_value(row, first_idx, inv_rms),
|
|
123
|
+
normalized_value(row, second_idx, inv_rms),
|
|
124
|
+
freqs_cos[freq_idx],
|
|
125
|
+
freqs_sin[freq_idx]
|
|
126
|
+
);
|
|
127
|
+
store_qk_value(row, first_idx, rotated.x);
|
|
128
|
+
store_qk_value(row, second_idx, rotated.y);
|
|
129
|
+
}
|
|
130
|
+
|
|
131
|
+
fn copy_v(v_group: u32, thread_idx: u32) {
|
|
132
|
+
let v_idx = v_group * WORKGROUP_SIZE + thread_idx;
|
|
133
|
+
if (v_idx >= u.total_v) {
|
|
134
|
+
return;
|
|
135
|
+
}
|
|
136
|
+
let token = v_idx / u.v_size;
|
|
137
|
+
let elem = v_idx % u.v_size;
|
|
138
|
+
let src_idx = token * qkv_size() + u.q_size + u.k_size + elem;
|
|
139
|
+
v_output[v_idx] = fused_qkv[src_idx];
|
|
140
|
+
}
|
|
141
|
+
|
|
142
|
+
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
143
|
+
fn main(
|
|
144
|
+
@builtin(local_invocation_id) local_id: vec3<u32>,
|
|
145
|
+
@builtin(workgroup_id) wg_id: vec3<u32>
|
|
146
|
+
) {
|
|
147
|
+
let group_idx = flat_group_index(wg_id);
|
|
148
|
+
let thread_idx = local_id.x;
|
|
149
|
+
|
|
150
|
+
if (group_idx >= u.qk_rows) {
|
|
151
|
+
copy_v(group_idx - u.qk_rows, thread_idx);
|
|
152
|
+
return;
|
|
153
|
+
}
|
|
154
|
+
|
|
155
|
+
let elements_per_thread = (u.head_dim + WORKGROUP_SIZE - 1u) / WORKGROUP_SIZE;
|
|
156
|
+
var local_sum_sq: f32 = 0.0;
|
|
157
|
+
for (var i: u32 = 0u; i < elements_per_thread; i = i + 1u) {
|
|
158
|
+
let idx = thread_idx * elements_per_thread + i;
|
|
159
|
+
if (idx < u.head_dim) {
|
|
160
|
+
let x = load_qk_value(group_idx, idx);
|
|
161
|
+
local_sum_sq = local_sum_sq + x * x;
|
|
162
|
+
}
|
|
163
|
+
}
|
|
164
|
+
|
|
165
|
+
shared_sum[thread_idx] = local_sum_sq;
|
|
166
|
+
workgroupBarrier();
|
|
167
|
+
|
|
168
|
+
for (var stride: u32 = WORKGROUP_SIZE / 2u; stride > 0u; stride = stride >> 1u) {
|
|
169
|
+
if (thread_idx < stride) {
|
|
170
|
+
shared_sum[thread_idx] = shared_sum[thread_idx] + shared_sum[thread_idx + stride];
|
|
171
|
+
}
|
|
172
|
+
workgroupBarrier();
|
|
173
|
+
}
|
|
174
|
+
|
|
175
|
+
let inv_rms = 1.0 / sqrt(shared_sum[0] / f32(u.head_dim) + u.eps);
|
|
176
|
+
workgroupBarrier();
|
|
177
|
+
|
|
178
|
+
let pairs_per_thread = (u.half_dim + WORKGROUP_SIZE - 1u) / WORKGROUP_SIZE;
|
|
179
|
+
for (var i: u32 = 0u; i < pairs_per_thread; i = i + 1u) {
|
|
180
|
+
let pair_idx = thread_idx * pairs_per_thread + i;
|
|
181
|
+
if (pair_idx < u.half_dim) {
|
|
182
|
+
store_rotated_qk_pair(group_idx, pair_idx, inv_rms);
|
|
183
|
+
}
|
|
184
|
+
}
|
|
185
|
+
}
|
|
@@ -0,0 +1,187 @@
|
|
|
1
|
+
enable f16;
|
|
2
|
+
|
|
3
|
+
override WORKGROUP_SIZE: u32 = 256u;
|
|
4
|
+
const MAX_WORKGROUP_SIZE: u32 = 256u;
|
|
5
|
+
|
|
6
|
+
override RMS_NORM_OFFSET: bool = false;
|
|
7
|
+
override Q_WEIGHT_IS_F16: bool = false;
|
|
8
|
+
override K_WEIGHT_IS_F16: bool = false;
|
|
9
|
+
|
|
10
|
+
struct Uniforms {
|
|
11
|
+
num_tokens: u32,
|
|
12
|
+
q_size: u32,
|
|
13
|
+
k_size: u32,
|
|
14
|
+
v_size: u32,
|
|
15
|
+
num_heads: u32,
|
|
16
|
+
num_kv_heads: u32,
|
|
17
|
+
head_dim: u32,
|
|
18
|
+
workgroup_stride: u32,
|
|
19
|
+
qk_rows: u32,
|
|
20
|
+
total_v: u32,
|
|
21
|
+
start_pos: u32,
|
|
22
|
+
half_dim: u32,
|
|
23
|
+
eps: f32,
|
|
24
|
+
kv_dst_offset: u32,
|
|
25
|
+
_pad1: u32,
|
|
26
|
+
_pad2: u32,
|
|
27
|
+
}
|
|
28
|
+
|
|
29
|
+
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
30
|
+
@group(0) @binding(1) var<storage, read> fused_qkv: array<f32>;
|
|
31
|
+
@group(0) @binding(2) var<storage, read> q_weight: array<u32>;
|
|
32
|
+
@group(0) @binding(3) var<storage, read> k_weight: array<u32>;
|
|
33
|
+
@group(0) @binding(4) var<storage, read> freqs_cos: array<f32>;
|
|
34
|
+
@group(0) @binding(5) var<storage, read> freqs_sin: array<f32>;
|
|
35
|
+
@group(0) @binding(6) var<storage, read_write> q_output: array<f32>;
|
|
36
|
+
@group(0) @binding(7) var<storage, read_write> kv_cache_k: array<f16>;
|
|
37
|
+
@group(0) @binding(8) var<storage, read_write> kv_cache_v: array<f16>;
|
|
38
|
+
|
|
39
|
+
var<workgroup> shared_sum: array<f32, MAX_WORKGROUP_SIZE>;
|
|
40
|
+
|
|
41
|
+
fn flat_group_index(wg_id: vec3<u32>) -> u32 {
|
|
42
|
+
return wg_id.y * max(u.workgroup_stride, 1u) + wg_id.x;
|
|
43
|
+
}
|
|
44
|
+
|
|
45
|
+
fn q_rows() -> u32 {
|
|
46
|
+
return u.num_tokens * u.num_heads;
|
|
47
|
+
}
|
|
48
|
+
|
|
49
|
+
fn qkv_size() -> u32 {
|
|
50
|
+
return u.q_size + u.k_size + u.v_size;
|
|
51
|
+
}
|
|
52
|
+
|
|
53
|
+
fn apply_weight(w: f32) -> f32 {
|
|
54
|
+
if (RMS_NORM_OFFSET) {
|
|
55
|
+
return 1.0 + w;
|
|
56
|
+
}
|
|
57
|
+
return w;
|
|
58
|
+
}
|
|
59
|
+
|
|
60
|
+
fn token_for_row(row: u32) -> u32 {
|
|
61
|
+
if (row < q_rows()) {
|
|
62
|
+
return row / u.num_heads;
|
|
63
|
+
}
|
|
64
|
+
let k_row = row - q_rows();
|
|
65
|
+
return k_row / u.num_kv_heads;
|
|
66
|
+
}
|
|
67
|
+
|
|
68
|
+
fn load_qk_value(row: u32, idx: u32) -> f32 {
|
|
69
|
+
if (row < q_rows()) {
|
|
70
|
+
let token = row / u.num_heads;
|
|
71
|
+
let head = row % u.num_heads;
|
|
72
|
+
return fused_qkv[token * qkv_size() + head * u.head_dim + idx];
|
|
73
|
+
}
|
|
74
|
+
|
|
75
|
+
let k_row = row - q_rows();
|
|
76
|
+
let token = k_row / u.num_kv_heads;
|
|
77
|
+
let head = k_row % u.num_kv_heads;
|
|
78
|
+
return fused_qkv[token * qkv_size() + u.q_size + head * u.head_dim + idx];
|
|
79
|
+
}
|
|
80
|
+
|
|
81
|
+
fn load_weight(row: u32, idx: u32) -> f32 {
|
|
82
|
+
if (row < q_rows()) {
|
|
83
|
+
if (Q_WEIGHT_IS_F16) {
|
|
84
|
+
let packed = q_weight[idx >> 1u];
|
|
85
|
+
let pair = unpack2x16float(packed);
|
|
86
|
+
return select(pair.x, pair.y, (idx & 1u) == 1u);
|
|
87
|
+
}
|
|
88
|
+
return bitcast<f32>(q_weight[idx]);
|
|
89
|
+
}
|
|
90
|
+
if (K_WEIGHT_IS_F16) {
|
|
91
|
+
let packed = k_weight[idx >> 1u];
|
|
92
|
+
let pair = unpack2x16float(packed);
|
|
93
|
+
return select(pair.x, pair.y, (idx & 1u) == 1u);
|
|
94
|
+
}
|
|
95
|
+
return bitcast<f32>(k_weight[idx]);
|
|
96
|
+
}
|
|
97
|
+
|
|
98
|
+
fn store_qk_value(row: u32, idx: u32, value: f32) {
|
|
99
|
+
if (row < q_rows()) {
|
|
100
|
+
q_output[row * u.head_dim + idx] = value;
|
|
101
|
+
} else {
|
|
102
|
+
let k_row = row - q_rows();
|
|
103
|
+
kv_cache_k[u.kv_dst_offset + k_row * u.head_dim + idx] = f16(value);
|
|
104
|
+
}
|
|
105
|
+
}
|
|
106
|
+
|
|
107
|
+
fn normalized_value(row: u32, idx: u32, inv_rms: f32) -> f32 {
|
|
108
|
+
return load_qk_value(row, idx) * inv_rms * apply_weight(load_weight(row, idx));
|
|
109
|
+
}
|
|
110
|
+
|
|
111
|
+
fn rotate_pair(x0: f32, x1: f32, cos_val: f32, sin_val: f32) -> vec2<f32> {
|
|
112
|
+
return vec2<f32>(
|
|
113
|
+
x0 * cos_val - x1 * sin_val,
|
|
114
|
+
x0 * sin_val + x1 * cos_val
|
|
115
|
+
);
|
|
116
|
+
}
|
|
117
|
+
|
|
118
|
+
fn store_rotated_qk_pair(row: u32, pair_idx: u32, inv_rms: f32) {
|
|
119
|
+
let first_idx = pair_idx;
|
|
120
|
+
let second_idx = pair_idx + u.half_dim;
|
|
121
|
+
let token = token_for_row(row);
|
|
122
|
+
let freq_idx = (u.start_pos + token) * u.half_dim + pair_idx;
|
|
123
|
+
let rotated = rotate_pair(
|
|
124
|
+
normalized_value(row, first_idx, inv_rms),
|
|
125
|
+
normalized_value(row, second_idx, inv_rms),
|
|
126
|
+
freqs_cos[freq_idx],
|
|
127
|
+
freqs_sin[freq_idx]
|
|
128
|
+
);
|
|
129
|
+
store_qk_value(row, first_idx, rotated.x);
|
|
130
|
+
store_qk_value(row, second_idx, rotated.y);
|
|
131
|
+
}
|
|
132
|
+
|
|
133
|
+
fn copy_v(v_group: u32, thread_idx: u32) {
|
|
134
|
+
let v_idx = v_group * WORKGROUP_SIZE + thread_idx;
|
|
135
|
+
if (v_idx >= u.total_v) {
|
|
136
|
+
return;
|
|
137
|
+
}
|
|
138
|
+
let token = v_idx / u.v_size;
|
|
139
|
+
let elem = v_idx % u.v_size;
|
|
140
|
+
let src_idx = token * qkv_size() + u.q_size + u.k_size + elem;
|
|
141
|
+
kv_cache_v[u.kv_dst_offset + v_idx] = f16(fused_qkv[src_idx]);
|
|
142
|
+
}
|
|
143
|
+
|
|
144
|
+
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
145
|
+
fn main(
|
|
146
|
+
@builtin(local_invocation_id) local_id: vec3<u32>,
|
|
147
|
+
@builtin(workgroup_id) wg_id: vec3<u32>
|
|
148
|
+
) {
|
|
149
|
+
let group_idx = flat_group_index(wg_id);
|
|
150
|
+
let thread_idx = local_id.x;
|
|
151
|
+
|
|
152
|
+
if (group_idx >= u.qk_rows) {
|
|
153
|
+
copy_v(group_idx - u.qk_rows, thread_idx);
|
|
154
|
+
return;
|
|
155
|
+
}
|
|
156
|
+
|
|
157
|
+
let elements_per_thread = (u.head_dim + WORKGROUP_SIZE - 1u) / WORKGROUP_SIZE;
|
|
158
|
+
var local_sum_sq: f32 = 0.0;
|
|
159
|
+
for (var i: u32 = 0u; i < elements_per_thread; i = i + 1u) {
|
|
160
|
+
let idx = thread_idx * elements_per_thread + i;
|
|
161
|
+
if (idx < u.head_dim) {
|
|
162
|
+
let x = load_qk_value(group_idx, idx);
|
|
163
|
+
local_sum_sq = local_sum_sq + x * x;
|
|
164
|
+
}
|
|
165
|
+
}
|
|
166
|
+
|
|
167
|
+
shared_sum[thread_idx] = local_sum_sq;
|
|
168
|
+
workgroupBarrier();
|
|
169
|
+
|
|
170
|
+
for (var stride: u32 = WORKGROUP_SIZE / 2u; stride > 0u; stride = stride >> 1u) {
|
|
171
|
+
if (thread_idx < stride) {
|
|
172
|
+
shared_sum[thread_idx] = shared_sum[thread_idx] + shared_sum[thread_idx + stride];
|
|
173
|
+
}
|
|
174
|
+
workgroupBarrier();
|
|
175
|
+
}
|
|
176
|
+
|
|
177
|
+
let inv_rms = 1.0 / sqrt(shared_sum[0] / f32(u.head_dim) + u.eps);
|
|
178
|
+
workgroupBarrier();
|
|
179
|
+
|
|
180
|
+
let pairs_per_thread = (u.half_dim + WORKGROUP_SIZE - 1u) / WORKGROUP_SIZE;
|
|
181
|
+
for (var i: u32 = 0u; i < pairs_per_thread; i = i + 1u) {
|
|
182
|
+
let pair_idx = thread_idx * pairs_per_thread + i;
|
|
183
|
+
if (pair_idx < u.half_dim) {
|
|
184
|
+
store_rotated_qk_pair(group_idx, pair_idx, inv_rms);
|
|
185
|
+
}
|
|
186
|
+
}
|
|
187
|
+
}
|
|
@@ -1,7 +1,7 @@
|
|
|
1
1
|
|
|
2
2
|
|
|
3
3
|
import { getDevice } from '../device.js';
|
|
4
|
-
import { getUniformCache } from '../uniform-cache.js';
|
|
4
|
+
import { getUniformCache, toUniformArrayBuffer } from '../uniform-cache.js';
|
|
5
5
|
|
|
6
6
|
// ============================================================================
|
|
7
7
|
// Uniform Buffer Creation
|
|
@@ -78,10 +78,7 @@ export function createUniformBufferFromData(
|
|
|
78
78
|
return recorder.createUniformBuffer(data, label);
|
|
79
79
|
}
|
|
80
80
|
|
|
81
|
-
|
|
82
|
-
const arrayBuffer = data instanceof ArrayBuffer
|
|
83
|
-
? data
|
|
84
|
-
: data.buffer.slice(data.byteOffset, data.byteOffset + data.byteLength);
|
|
81
|
+
const arrayBuffer = toUniformArrayBuffer(data);
|
|
85
82
|
|
|
86
83
|
// Use cache by default for non-recorder paths
|
|
87
84
|
const useCache = options?.useCache ?? true;
|
|
@@ -45,6 +45,7 @@ export {
|
|
|
45
45
|
getOrCreatePipelineLayout,
|
|
46
46
|
getCachedPipeline,
|
|
47
47
|
getPipelineFast,
|
|
48
|
+
getPipelineBindGroupLayout,
|
|
48
49
|
createPipeline,
|
|
49
50
|
clearPipelineCaches,
|
|
50
51
|
getPipelineCacheStats,
|
|
@@ -86,6 +87,19 @@ export {
|
|
|
86
87
|
// Combined Cache Management
|
|
87
88
|
// ============================================================================
|
|
88
89
|
|
|
90
|
+
/** Whether debug-only kernel record-stage timing is enabled for this process. */
|
|
91
|
+
export declare const RECORD_STAGE_DEBUG_ENABLED: boolean;
|
|
92
|
+
|
|
93
|
+
/** Debug-only kernel timing recorder enabled by DOPPLER_DBG_RECORD. */
|
|
94
|
+
export declare function __dbgRecord(
|
|
95
|
+
op: string,
|
|
96
|
+
variant: string,
|
|
97
|
+
pipelineMs: number,
|
|
98
|
+
prepMs: number,
|
|
99
|
+
bgMs: number,
|
|
100
|
+
dispatchMs: number
|
|
101
|
+
): void;
|
|
102
|
+
|
|
89
103
|
/**
|
|
90
104
|
* Unified kernel dispatch helper. Resolves the kernel config and pipeline
|
|
91
105
|
* for `opName`/`variant`, creates a uniform buffer from `uniforms`, and
|
|
@@ -99,7 +113,8 @@ export declare function unifiedKernelWrapper(
|
|
|
99
113
|
uniforms: Record<string, number>,
|
|
100
114
|
workgroups: number | [number, number, number],
|
|
101
115
|
constants?: Record<string, number> | null,
|
|
102
|
-
extraBindings?: unknown[] | null
|
|
116
|
+
extraBindings?: unknown[] | null,
|
|
117
|
+
dispatchLabel?: string | null
|
|
103
118
|
): Promise<void>;
|
|
104
119
|
|
|
105
120
|
/**
|
package/src/gpu/kernels/utils.js
CHANGED
|
@@ -1,5 +1,7 @@
|
|
|
1
1
|
|
|
2
2
|
|
|
3
|
+
export const RECORD_STAGE_DEBUG_ENABLED = typeof process !== 'undefined' && process.env?.DOPPLER_DBG_RECORD === '1';
|
|
4
|
+
|
|
3
5
|
const __DBG_REC = { byOp: new Map(), totalCount: 0, total: { pipeline: 0, prep: 0, bg: 0, dispatch: 0 }, firstCallByOp: new Map() };
|
|
4
6
|
export function __dbgRecord(op, variant, pipelineMs, prepMs, bgMs, dispatchMs) {
|
|
5
7
|
const key = `${op}/${variant}`;
|
|
@@ -17,7 +19,7 @@ export function __dbgRecord(op, variant, pipelineMs, prepMs, bgMs, dispatchMs) {
|
|
|
17
19
|
__DBG_REC.total.bg += bgMs;
|
|
18
20
|
__DBG_REC.total.dispatch += dispatchMs;
|
|
19
21
|
}
|
|
20
|
-
if (
|
|
22
|
+
if (RECORD_STAGE_DEBUG_ENABLED) {
|
|
21
23
|
process.on('exit', () => {
|
|
22
24
|
const t = __DBG_REC.total;
|
|
23
25
|
const sum = t.pipeline + t.prep + t.bg + t.dispatch;
|
|
@@ -63,6 +65,7 @@ export {
|
|
|
63
65
|
getOrCreatePipelineLayout,
|
|
64
66
|
getCachedPipeline,
|
|
65
67
|
getPipelineFast,
|
|
68
|
+
getPipelineBindGroupLayout,
|
|
66
69
|
createPipeline,
|
|
67
70
|
clearPipelineCaches,
|
|
68
71
|
getPipelineCacheStats,
|
|
@@ -103,20 +106,46 @@ export {
|
|
|
103
106
|
// ============================================================================
|
|
104
107
|
|
|
105
108
|
import { getKernelConfig } from './kernel-configs.js';
|
|
106
|
-
import { getPipelineFast } from './pipeline-cache.js';
|
|
109
|
+
import { getCachedPipeline, getPipelineFast, getPipelineBindGroupLayout } from './pipeline-cache.js';
|
|
107
110
|
import { getDevice } from '../device.js';
|
|
108
111
|
import { dispatchKernel, dispatchIndirect, recordDispatchIndirect } from './dispatch.js';
|
|
109
112
|
import { createUniformBufferWithView as createUniformBuffer } from './uniform-utils.js';
|
|
110
113
|
import { getUniformByteLength } from './uniform-utils.js';
|
|
111
114
|
import { writeUniformsFromObject } from './uniform-utils.js';
|
|
112
115
|
|
|
113
|
-
|
|
114
|
-
|
|
116
|
+
const dataBindingsByConfig = new WeakMap();
|
|
117
|
+
|
|
118
|
+
function getDataBindings(config) {
|
|
119
|
+
let cached = dataBindingsByConfig.get(config);
|
|
120
|
+
if (cached) {
|
|
121
|
+
return cached;
|
|
122
|
+
}
|
|
123
|
+
cached = config.bindings
|
|
124
|
+
.filter(b => b.type !== 'uniform')
|
|
125
|
+
.slice()
|
|
126
|
+
.sort((a, b) => a.index - b.index);
|
|
127
|
+
dataBindingsByConfig.set(config, cached);
|
|
128
|
+
return cached;
|
|
129
|
+
}
|
|
130
|
+
|
|
131
|
+
export async function unifiedKernelWrapper(
|
|
132
|
+
opName,
|
|
133
|
+
target,
|
|
134
|
+
variant,
|
|
135
|
+
bindings,
|
|
136
|
+
uniforms,
|
|
137
|
+
workgroups,
|
|
138
|
+
constants = null,
|
|
139
|
+
extraBindings = null,
|
|
140
|
+
dispatchLabel = null
|
|
141
|
+
) {
|
|
142
|
+
const __dbg = RECORD_STAGE_DEBUG_ENABLED;
|
|
115
143
|
const __t0 = __dbg ? performance.now() : 0;
|
|
116
144
|
const device = target?.device || getDevice();
|
|
117
145
|
const recorder = target && typeof target.beginComputePass === 'function' ? target : null;
|
|
118
146
|
const config = getKernelConfig(opName, variant);
|
|
119
|
-
const pipeline =
|
|
147
|
+
const pipeline = getCachedPipeline(opName, variant, constants)
|
|
148
|
+
?? await getPipelineFast(opName, variant, null, constants);
|
|
120
149
|
const __tPipeline = __dbg ? performance.now() : 0;
|
|
121
150
|
|
|
122
151
|
const uniformBuffer = createUniformBuffer(
|
|
@@ -131,10 +160,7 @@ export async function unifiedKernelWrapper(opName, target, variant, bindings, un
|
|
|
131
160
|
{ binding: 0, resource: { buffer: uniformBuffer } }
|
|
132
161
|
];
|
|
133
162
|
|
|
134
|
-
const dataBindings = config
|
|
135
|
-
.filter(b => b.type !== 'uniform')
|
|
136
|
-
.slice()
|
|
137
|
-
.sort((a, b) => a.index - b.index);
|
|
163
|
+
const dataBindings = getDataBindings(config);
|
|
138
164
|
|
|
139
165
|
if (bindings.length !== dataBindings.length) {
|
|
140
166
|
throw new Error(
|
|
@@ -189,20 +215,23 @@ export async function unifiedKernelWrapper(opName, target, variant, bindings, un
|
|
|
189
215
|
const __tBgStart = __dbg ? performance.now() : 0;
|
|
190
216
|
const bindGroup = device.createBindGroup({
|
|
191
217
|
label: `${opName}_bind_group`,
|
|
192
|
-
layout: pipeline
|
|
218
|
+
layout: getPipelineBindGroupLayout(pipeline, 0),
|
|
193
219
|
entries: bindGroupEntries,
|
|
194
220
|
});
|
|
195
221
|
const __tBg = __dbg ? performance.now() : 0;
|
|
196
222
|
|
|
223
|
+
const label = typeof dispatchLabel === 'string' && dispatchLabel.length > 0
|
|
224
|
+
? dispatchLabel
|
|
225
|
+
: opName;
|
|
197
226
|
if (workgroups && typeof workgroups === 'object' && workgroups.indirectBuffer) {
|
|
198
227
|
const indirectOffset = workgroups.indirectOffset ?? 0;
|
|
199
228
|
if (recorder) {
|
|
200
|
-
recordDispatchIndirect(recorder, pipeline, bindGroup, workgroups.indirectBuffer, indirectOffset,
|
|
229
|
+
recordDispatchIndirect(recorder, pipeline, bindGroup, workgroups.indirectBuffer, indirectOffset, label);
|
|
201
230
|
} else {
|
|
202
|
-
dispatchIndirect(device, pipeline, bindGroup, workgroups.indirectBuffer, indirectOffset,
|
|
231
|
+
dispatchIndirect(device, pipeline, bindGroup, workgroups.indirectBuffer, indirectOffset, label);
|
|
203
232
|
}
|
|
204
233
|
} else {
|
|
205
|
-
dispatchKernel(target, pipeline, bindGroup, workgroups,
|
|
234
|
+
dispatchKernel(target, pipeline, bindGroup, workgroups, label);
|
|
206
235
|
}
|
|
207
236
|
if (__dbg) {
|
|
208
237
|
const __tEnd = performance.now();
|