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 +1,49 @@
|
|
|
1
|
+
// scatter_add_dynamic_f16_weights_expert_scale.wgsl
|
|
2
|
+
|
|
3
|
+
enable f16;
|
|
4
|
+
|
|
5
|
+
override WORKGROUP_SIZE: u32 = 256u;
|
|
6
|
+
|
|
7
|
+
struct DynamicUniforms {
|
|
8
|
+
num_tokens: u32,
|
|
9
|
+
hidden_size: u32,
|
|
10
|
+
top_k: u32,
|
|
11
|
+
_pad: u32,
|
|
12
|
+
}
|
|
13
|
+
|
|
14
|
+
@group(0) @binding(0) var<uniform> u_dyn: DynamicUniforms;
|
|
15
|
+
@group(0) @binding(1) var<storage, read> expert_outputs_flat: array<f16>;
|
|
16
|
+
@group(0) @binding(2) var<storage, read> routing_indices: array<u32>;
|
|
17
|
+
@group(0) @binding(3) var<storage, read> routing_weights: array<f16>;
|
|
18
|
+
@group(0) @binding(4) var<storage, read> token_offsets: array<u32>;
|
|
19
|
+
@group(0) @binding(5) var<storage, read> expert_scales: array<f32>;
|
|
20
|
+
@group(0) @binding(6) var<storage, read_write> output_dyn: array<f16>;
|
|
21
|
+
|
|
22
|
+
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
23
|
+
fn scatter_add_dynamic(@builtin(global_invocation_id) gid: vec3<u32>) {
|
|
24
|
+
let tid = gid.x;
|
|
25
|
+
let total_elements = u_dyn.num_tokens * u_dyn.hidden_size;
|
|
26
|
+
|
|
27
|
+
if (tid >= total_elements) {
|
|
28
|
+
return;
|
|
29
|
+
}
|
|
30
|
+
|
|
31
|
+
let token_idx = tid / u_dyn.hidden_size;
|
|
32
|
+
let dim_idx = tid % u_dyn.hidden_size;
|
|
33
|
+
let top_k = u_dyn.top_k;
|
|
34
|
+
let hidden_size = u_dyn.hidden_size;
|
|
35
|
+
|
|
36
|
+
var sum: f16 = f16(0.0);
|
|
37
|
+
let routing_base = token_idx * top_k;
|
|
38
|
+
|
|
39
|
+
for (var k: u32 = 0u; k < top_k; k = k + 1u) {
|
|
40
|
+
let expert_idx = routing_indices[routing_base + k];
|
|
41
|
+
let weight = routing_weights[routing_base + k] * f16(expert_scales[expert_idx]);
|
|
42
|
+
let data_offset = token_offsets[routing_base + k];
|
|
43
|
+
let expert_data_idx = data_offset * hidden_size + dim_idx;
|
|
44
|
+
|
|
45
|
+
sum = sum + weight * expert_outputs_flat[expert_data_idx];
|
|
46
|
+
}
|
|
47
|
+
|
|
48
|
+
output_dyn[tid] = sum;
|
|
49
|
+
}
|
|
@@ -0,0 +1,45 @@
|
|
|
1
|
+
// scatter_add_routes_f16_weights_expert_scale.wgsl
|
|
2
|
+
|
|
3
|
+
enable f16;
|
|
4
|
+
|
|
5
|
+
override WORKGROUP_SIZE: u32 = 256u;
|
|
6
|
+
|
|
7
|
+
struct RouteUniforms {
|
|
8
|
+
num_tokens: u32,
|
|
9
|
+
hidden_size: u32,
|
|
10
|
+
top_k: u32,
|
|
11
|
+
_pad0: u32,
|
|
12
|
+
}
|
|
13
|
+
|
|
14
|
+
@group(0) @binding(0) var<uniform> u: RouteUniforms;
|
|
15
|
+
@group(0) @binding(1) var<storage, read> route_outputs: array<f16>;
|
|
16
|
+
@group(0) @binding(2) var<storage, read> routing_indices: array<u32>;
|
|
17
|
+
@group(0) @binding(3) var<storage, read> routing_weights: array<f16>;
|
|
18
|
+
@group(0) @binding(4) var<storage, read> expert_scales: array<f32>;
|
|
19
|
+
@group(0) @binding(5) var<storage, read_write> output: array<f16>;
|
|
20
|
+
|
|
21
|
+
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
22
|
+
fn scatter_add_routes(@builtin(global_invocation_id) gid: vec3<u32>) {
|
|
23
|
+
let total_elements = u.num_tokens * u.hidden_size;
|
|
24
|
+
let base_tid = gid.x * 4u;
|
|
25
|
+
|
|
26
|
+
for (var lane: u32 = 0u; lane < 4u; lane = lane + 1u) {
|
|
27
|
+
let tid = base_tid + lane;
|
|
28
|
+
if (tid < total_elements) {
|
|
29
|
+
let token_idx = tid / u.hidden_size;
|
|
30
|
+
let dim_idx = tid % u.hidden_size;
|
|
31
|
+
let route_base = token_idx * u.top_k;
|
|
32
|
+
|
|
33
|
+
var sum: f16 = f16(0.0);
|
|
34
|
+
for (var k: u32 = 0u; k < u.top_k; k = k + 1u) {
|
|
35
|
+
let route_idx = route_base + k;
|
|
36
|
+
let expert_idx = routing_indices[route_idx];
|
|
37
|
+
let weight = routing_weights[route_idx] * f16(expert_scales[expert_idx]);
|
|
38
|
+
let route_output_idx = route_idx * u.hidden_size + dim_idx;
|
|
39
|
+
sum = sum + weight * route_outputs[route_output_idx];
|
|
40
|
+
}
|
|
41
|
+
|
|
42
|
+
output[tid] = sum;
|
|
43
|
+
}
|
|
44
|
+
}
|
|
45
|
+
}
|
|
@@ -66,27 +66,7 @@ function getKernelBasePath() {
|
|
|
66
66
|
return override.replace(/\/+$/, '');
|
|
67
67
|
}
|
|
68
68
|
|
|
69
|
-
|
|
70
|
-
if (typeof location === 'undefined') {
|
|
71
|
-
return new URL('.', import.meta.url).toString().replace(/\/$/, '');
|
|
72
|
-
}
|
|
73
|
-
|
|
74
|
-
// Check common deployed Doppler paths.
|
|
75
|
-
if (typeof location !== 'undefined') {
|
|
76
|
-
const path = location.pathname || '';
|
|
77
|
-
if (
|
|
78
|
-
path === '/d' ||
|
|
79
|
-
path.startsWith('/d/') ||
|
|
80
|
-
path === '/doppler' ||
|
|
81
|
-
path.startsWith('/doppler/') ||
|
|
82
|
-
path === '/dr' ||
|
|
83
|
-
path.startsWith('/dr/') ||
|
|
84
|
-
location.host.includes('replo')
|
|
85
|
-
) {
|
|
86
|
-
return '/doppler/src/gpu/kernels';
|
|
87
|
-
}
|
|
88
|
-
}
|
|
89
|
-
return '/src/gpu/kernels';
|
|
69
|
+
return new URL('.', import.meta.url).toString().replace(/\/$/, '');
|
|
90
70
|
}
|
|
91
71
|
|
|
92
72
|
const KERNEL_BASE_PATH = getKernelBasePath();
|
|
@@ -0,0 +1,28 @@
|
|
|
1
|
+
import type { Tensor } from '../tensor.js';
|
|
2
|
+
import type { SplitWeightBuffer, WeightBuffer } from '../weight-buffer.js';
|
|
3
|
+
import type { OutputBufferOptions } from './types.js';
|
|
4
|
+
|
|
5
|
+
export interface SoftEmbeddingSplitOptions extends OutputBufferOptions {}
|
|
6
|
+
|
|
7
|
+
export interface SoftEmbeddingLogitsOptions {
|
|
8
|
+
temperature?: number;
|
|
9
|
+
chunkRows?: number;
|
|
10
|
+
}
|
|
11
|
+
|
|
12
|
+
export declare function runSoftEmbeddingSplitF16(
|
|
13
|
+
softmaxTensor: Tensor,
|
|
14
|
+
splitEmbedding: SplitWeightBuffer,
|
|
15
|
+
numTokens: number,
|
|
16
|
+
hiddenSize: number,
|
|
17
|
+
vocabSize: number,
|
|
18
|
+
options?: SoftEmbeddingSplitOptions
|
|
19
|
+
): Promise<Tensor>;
|
|
20
|
+
|
|
21
|
+
export declare function runSoftEmbeddingLogitsF16(
|
|
22
|
+
logitsTensor: Tensor,
|
|
23
|
+
embedding: WeightBuffer,
|
|
24
|
+
numTokens: number,
|
|
25
|
+
hiddenSize: number,
|
|
26
|
+
vocabSize: number,
|
|
27
|
+
options?: SoftEmbeddingLogitsOptions
|
|
28
|
+
): Promise<Tensor>;
|
|
@@ -0,0 +1,447 @@
|
|
|
1
|
+
import { getDevice } from '../device.js';
|
|
2
|
+
import { createTensor } from '../tensor.js';
|
|
3
|
+
import { isSplitWeightBuffer, isWeightBuffer, getWeightDtype, getLayout } from '../weight-buffer.js';
|
|
4
|
+
import { acquireBuffer, releaseBuffer } from '../../memory/buffer-pool.js';
|
|
5
|
+
import { createPipeline, createUniformBufferWithView, createBindGroupWithValidation } from './utils.js';
|
|
6
|
+
import { dispatchKernel } from './dispatch.js';
|
|
7
|
+
import { runMatmul } from './matmul.js';
|
|
8
|
+
import { runResidualAdd } from './residual.js';
|
|
9
|
+
|
|
10
|
+
const UNIFORM_SIZE = 32;
|
|
11
|
+
const WORKGROUP_X = 8;
|
|
12
|
+
const WORKGROUP_Y = 8;
|
|
13
|
+
const LOGITS_WORKGROUP_SIZE = 256;
|
|
14
|
+
const DEFAULT_LOGITS_CHUNK_ROWS = 32768;
|
|
15
|
+
const STORAGE_OFFSET_ALIGNMENT = 256;
|
|
16
|
+
|
|
17
|
+
function validatePositiveInt(value, label) {
|
|
18
|
+
if (!Number.isInteger(value) || value <= 0) {
|
|
19
|
+
throw new Error(`[SoftEmbedding] ${label} must be a positive integer; got ${String(value)}.`);
|
|
20
|
+
}
|
|
21
|
+
}
|
|
22
|
+
|
|
23
|
+
function validatePositiveNumber(value, label) {
|
|
24
|
+
if (typeof value !== 'number' || !Number.isFinite(value) || value <= 0) {
|
|
25
|
+
throw new Error(`[SoftEmbedding] ${label} must be a positive finite number; got ${String(value)}.`);
|
|
26
|
+
}
|
|
27
|
+
}
|
|
28
|
+
|
|
29
|
+
function gcd(a, b) {
|
|
30
|
+
let x = Math.abs(a);
|
|
31
|
+
let y = Math.abs(b);
|
|
32
|
+
while (y !== 0) {
|
|
33
|
+
const next = x % y;
|
|
34
|
+
x = y;
|
|
35
|
+
y = next;
|
|
36
|
+
}
|
|
37
|
+
return x;
|
|
38
|
+
}
|
|
39
|
+
|
|
40
|
+
function alignChunkRowsForWeightOffset(chunkRows, hiddenSize, vocabSize) {
|
|
41
|
+
if (chunkRows >= vocabSize) {
|
|
42
|
+
return vocabSize;
|
|
43
|
+
}
|
|
44
|
+
const rowBytes = hiddenSize * Uint16Array.BYTES_PER_ELEMENT;
|
|
45
|
+
const alignmentRows = STORAGE_OFFSET_ALIGNMENT / gcd(rowBytes, STORAGE_OFFSET_ALIGNMENT);
|
|
46
|
+
const alignedRows = Math.floor(chunkRows / alignmentRows) * alignmentRows;
|
|
47
|
+
return Math.min(vocabSize, Math.max(alignmentRows, alignedRows));
|
|
48
|
+
}
|
|
49
|
+
|
|
50
|
+
function resolveLogitsChunkRows(device, numTokens, hiddenSize, vocabSize, requestedRows) {
|
|
51
|
+
const requested = Number.isInteger(requestedRows) && requestedRows > 0
|
|
52
|
+
? requestedRows
|
|
53
|
+
: DEFAULT_LOGITS_CHUNK_ROWS;
|
|
54
|
+
const maxBinding = Math.min(
|
|
55
|
+
device.limits?.maxStorageBufferBindingSize ?? Number.POSITIVE_INFINITY,
|
|
56
|
+
device.limits?.maxBufferSize ?? Number.POSITIVE_INFINITY
|
|
57
|
+
);
|
|
58
|
+
const maxByProbability = Number.isFinite(maxBinding)
|
|
59
|
+
? Math.floor(maxBinding / (numTokens * Float32Array.BYTES_PER_ELEMENT))
|
|
60
|
+
: requested;
|
|
61
|
+
const maxByEmbedding = Number.isFinite(maxBinding)
|
|
62
|
+
? Math.floor(maxBinding / (hiddenSize * Uint16Array.BYTES_PER_ELEMENT))
|
|
63
|
+
: requested;
|
|
64
|
+
const bounded = Math.max(1, Math.min(requested, maxByProbability, maxByEmbedding, vocabSize));
|
|
65
|
+
return alignChunkRowsForWeightOffset(bounded, hiddenSize, vocabSize);
|
|
66
|
+
}
|
|
67
|
+
|
|
68
|
+
function createSectionUniformBuffer(device, section, sectionIndex, numTokens, hiddenSize, vocabSize) {
|
|
69
|
+
return createUniformBufferWithView(
|
|
70
|
+
'soft_embedding_split_uniforms',
|
|
71
|
+
UNIFORM_SIZE,
|
|
72
|
+
(view) => {
|
|
73
|
+
view.setUint32(0, numTokens, true);
|
|
74
|
+
view.setUint32(4, hiddenSize, true);
|
|
75
|
+
view.setUint32(8, vocabSize, true);
|
|
76
|
+
view.setUint32(12, section.rowStart, true);
|
|
77
|
+
view.setUint32(16, section.rowCount, true);
|
|
78
|
+
view.setUint32(20, sectionIndex === 0 ? 0 : 1, true);
|
|
79
|
+
view.setUint32(24, 0, true);
|
|
80
|
+
view.setUint32(28, 0, true);
|
|
81
|
+
},
|
|
82
|
+
null,
|
|
83
|
+
device
|
|
84
|
+
);
|
|
85
|
+
}
|
|
86
|
+
|
|
87
|
+
function createLogitsUniformBuffer(device, numTokens, hiddenSize, vocabSize, rowStart, rowCount, temperature) {
|
|
88
|
+
return createUniformBufferWithView(
|
|
89
|
+
'soft_embedding_logits_uniforms',
|
|
90
|
+
UNIFORM_SIZE,
|
|
91
|
+
(view) => {
|
|
92
|
+
view.setUint32(0, numTokens, true);
|
|
93
|
+
view.setUint32(4, hiddenSize, true);
|
|
94
|
+
view.setUint32(8, vocabSize, true);
|
|
95
|
+
view.setUint32(12, rowStart, true);
|
|
96
|
+
view.setUint32(16, rowCount, true);
|
|
97
|
+
view.setFloat32(20, temperature, true);
|
|
98
|
+
view.setUint32(24, 0, true);
|
|
99
|
+
view.setUint32(28, 0, true);
|
|
100
|
+
},
|
|
101
|
+
null,
|
|
102
|
+
device
|
|
103
|
+
);
|
|
104
|
+
}
|
|
105
|
+
|
|
106
|
+
function validateSplitSections(splitEmbedding, vocabSize) {
|
|
107
|
+
const sections = splitEmbedding.sections;
|
|
108
|
+
if (!Array.isArray(sections) || sections.length === 0) {
|
|
109
|
+
throw new Error('[SoftEmbedding] split embedding requires at least one section.');
|
|
110
|
+
}
|
|
111
|
+
let nextRowStart = 0;
|
|
112
|
+
for (let sectionIndex = 0; sectionIndex < sections.length; sectionIndex += 1) {
|
|
113
|
+
const section = sections[sectionIndex];
|
|
114
|
+
if (!section?.buffer) {
|
|
115
|
+
throw new Error(`[SoftEmbedding] split section ${sectionIndex} is missing a GPU buffer.`);
|
|
116
|
+
}
|
|
117
|
+
if (section.rowStart !== nextRowStart) {
|
|
118
|
+
throw new Error(
|
|
119
|
+
`[SoftEmbedding] split section ${sectionIndex} rowStart=${section.rowStart} ` +
|
|
120
|
+
`is not contiguous from row ${nextRowStart}.`
|
|
121
|
+
);
|
|
122
|
+
}
|
|
123
|
+
if (!Number.isInteger(section.rowCount) || section.rowCount <= 0) {
|
|
124
|
+
throw new Error(`[SoftEmbedding] split section ${sectionIndex} has invalid rowCount.`);
|
|
125
|
+
}
|
|
126
|
+
nextRowStart += section.rowCount;
|
|
127
|
+
}
|
|
128
|
+
if (nextRowStart !== vocabSize) {
|
|
129
|
+
throw new Error(
|
|
130
|
+
`[SoftEmbedding] split embedding exposes ${nextRowStart} rows but vocabSize=${vocabSize}.`
|
|
131
|
+
);
|
|
132
|
+
}
|
|
133
|
+
}
|
|
134
|
+
|
|
135
|
+
async function dispatchSplitSection({
|
|
136
|
+
device,
|
|
137
|
+
pipeline,
|
|
138
|
+
softmaxTensor,
|
|
139
|
+
section,
|
|
140
|
+
sectionIndex,
|
|
141
|
+
output,
|
|
142
|
+
numTokens,
|
|
143
|
+
hiddenSize,
|
|
144
|
+
vocabSize,
|
|
145
|
+
}) {
|
|
146
|
+
const uniformBuffer = createSectionUniformBuffer(
|
|
147
|
+
device,
|
|
148
|
+
section,
|
|
149
|
+
sectionIndex,
|
|
150
|
+
numTokens,
|
|
151
|
+
hiddenSize,
|
|
152
|
+
vocabSize
|
|
153
|
+
);
|
|
154
|
+
|
|
155
|
+
try {
|
|
156
|
+
const bindGroup = await createBindGroupWithValidation(device, {
|
|
157
|
+
label: 'soft_embedding_split_bind_group',
|
|
158
|
+
layout: pipeline.getBindGroupLayout(0),
|
|
159
|
+
entries: [
|
|
160
|
+
{ binding: 0, resource: { buffer: uniformBuffer } },
|
|
161
|
+
{ binding: 1, resource: { buffer: softmaxTensor.buffer } },
|
|
162
|
+
{ binding: 2, resource: { buffer: section.buffer } },
|
|
163
|
+
{ binding: 3, resource: { buffer: output } },
|
|
164
|
+
],
|
|
165
|
+
}, 'soft_embedding/split_f16');
|
|
166
|
+
|
|
167
|
+
dispatchKernel(
|
|
168
|
+
null,
|
|
169
|
+
pipeline,
|
|
170
|
+
bindGroup,
|
|
171
|
+
[Math.ceil(hiddenSize / WORKGROUP_X), Math.ceil(numTokens / WORKGROUP_Y), 1],
|
|
172
|
+
'soft_embedding_split_f16'
|
|
173
|
+
);
|
|
174
|
+
} finally {
|
|
175
|
+
uniformBuffer.destroy();
|
|
176
|
+
}
|
|
177
|
+
}
|
|
178
|
+
|
|
179
|
+
export async function runSoftEmbeddingSplitF16(softmaxTensor, splitEmbedding, numTokens, hiddenSize, vocabSize, options = {}) {
|
|
180
|
+
validatePositiveInt(numTokens, 'numTokens');
|
|
181
|
+
validatePositiveInt(hiddenSize, 'hiddenSize');
|
|
182
|
+
validatePositiveInt(vocabSize, 'vocabSize');
|
|
183
|
+
|
|
184
|
+
if (softmaxTensor?.dtype !== 'f32') {
|
|
185
|
+
throw new Error(`[SoftEmbedding] split f16 path requires f32 softmax input, got "${softmaxTensor?.dtype ?? 'missing'}".`);
|
|
186
|
+
}
|
|
187
|
+
if (!isSplitWeightBuffer(splitEmbedding)) {
|
|
188
|
+
throw new Error('[SoftEmbedding] split f16 path requires a SplitWeightBuffer embedding table.');
|
|
189
|
+
}
|
|
190
|
+
if (splitEmbedding.dtype !== 'f16' || splitEmbedding.layout !== 'row') {
|
|
191
|
+
throw new Error(
|
|
192
|
+
`[SoftEmbedding] split path supports row-major f16 embeddings only; ` +
|
|
193
|
+
`got dtype=${splitEmbedding.dtype}, layout=${splitEmbedding.layout}.`
|
|
194
|
+
);
|
|
195
|
+
}
|
|
196
|
+
validateSplitSections(splitEmbedding, vocabSize);
|
|
197
|
+
const shape = Array.isArray(splitEmbedding.shape) ? splitEmbedding.shape : null;
|
|
198
|
+
if (!shape || shape.length !== 2 || shape[0] !== vocabSize || shape[1] !== hiddenSize) {
|
|
199
|
+
throw new Error(
|
|
200
|
+
`[SoftEmbedding] split embedding shape mismatch: expected [${vocabSize}, ${hiddenSize}], ` +
|
|
201
|
+
`got ${shape ? `[${shape.join(', ')}]` : 'missing'}.`
|
|
202
|
+
);
|
|
203
|
+
}
|
|
204
|
+
|
|
205
|
+
const device = getDevice();
|
|
206
|
+
if (!device) {
|
|
207
|
+
throw new Error('[SoftEmbedding] GPU device not available.');
|
|
208
|
+
}
|
|
209
|
+
|
|
210
|
+
const outputBytes = numTokens * hiddenSize * Float32Array.BYTES_PER_ELEMENT;
|
|
211
|
+
const output = options.outputBuffer ?? acquireBuffer(outputBytes, undefined, 'soft_embedding_split_output');
|
|
212
|
+
const ownsOutput = options.outputBuffer ? null : output;
|
|
213
|
+
const pipeline = await createPipeline('soft_embedding', 'split_f16');
|
|
214
|
+
|
|
215
|
+
try {
|
|
216
|
+
for (let sectionIndex = 0; sectionIndex < splitEmbedding.sections.length; sectionIndex += 1) {
|
|
217
|
+
await dispatchSplitSection({
|
|
218
|
+
device,
|
|
219
|
+
pipeline,
|
|
220
|
+
softmaxTensor,
|
|
221
|
+
section: splitEmbedding.sections[sectionIndex],
|
|
222
|
+
sectionIndex,
|
|
223
|
+
output,
|
|
224
|
+
numTokens,
|
|
225
|
+
hiddenSize,
|
|
226
|
+
vocabSize,
|
|
227
|
+
});
|
|
228
|
+
}
|
|
229
|
+
} catch (error) {
|
|
230
|
+
if (ownsOutput) {
|
|
231
|
+
releaseBuffer(ownsOutput);
|
|
232
|
+
}
|
|
233
|
+
throw error;
|
|
234
|
+
}
|
|
235
|
+
|
|
236
|
+
return createTensor(output, 'f32', [numTokens, hiddenSize], 'soft_embedding_split_output');
|
|
237
|
+
}
|
|
238
|
+
|
|
239
|
+
async function dispatchLogitsNormStats({
|
|
240
|
+
device,
|
|
241
|
+
logitsTensor,
|
|
242
|
+
rowMaxBuffer,
|
|
243
|
+
rowSumBuffer,
|
|
244
|
+
numTokens,
|
|
245
|
+
hiddenSize,
|
|
246
|
+
vocabSize,
|
|
247
|
+
temperature,
|
|
248
|
+
}) {
|
|
249
|
+
const pipeline = await createPipeline('soft_embedding', 'logits_norm_stats');
|
|
250
|
+
const uniformBuffer = createLogitsUniformBuffer(
|
|
251
|
+
device,
|
|
252
|
+
numTokens,
|
|
253
|
+
hiddenSize,
|
|
254
|
+
vocabSize,
|
|
255
|
+
0,
|
|
256
|
+
vocabSize,
|
|
257
|
+
temperature
|
|
258
|
+
);
|
|
259
|
+
try {
|
|
260
|
+
const bindGroup = await createBindGroupWithValidation(device, {
|
|
261
|
+
label: 'soft_embedding_logits_norm_stats_bind_group',
|
|
262
|
+
layout: pipeline.getBindGroupLayout(0),
|
|
263
|
+
entries: [
|
|
264
|
+
{ binding: 0, resource: { buffer: uniformBuffer } },
|
|
265
|
+
{ binding: 1, resource: { buffer: logitsTensor.buffer } },
|
|
266
|
+
{ binding: 2, resource: { buffer: rowMaxBuffer } },
|
|
267
|
+
{ binding: 3, resource: { buffer: rowSumBuffer } },
|
|
268
|
+
],
|
|
269
|
+
}, 'soft_embedding/logits_norm_stats');
|
|
270
|
+
dispatchKernel(null, pipeline, bindGroup, numTokens, 'soft_embedding_logits_norm_stats');
|
|
271
|
+
} finally {
|
|
272
|
+
uniformBuffer.destroy();
|
|
273
|
+
}
|
|
274
|
+
}
|
|
275
|
+
|
|
276
|
+
async function dispatchProbabilityChunk({
|
|
277
|
+
device,
|
|
278
|
+
logitsTensor,
|
|
279
|
+
rowMaxBuffer,
|
|
280
|
+
rowSumBuffer,
|
|
281
|
+
probabilityBuffer,
|
|
282
|
+
numTokens,
|
|
283
|
+
hiddenSize,
|
|
284
|
+
vocabSize,
|
|
285
|
+
rowStart,
|
|
286
|
+
rowCount,
|
|
287
|
+
temperature,
|
|
288
|
+
}) {
|
|
289
|
+
const pipeline = await createPipeline('soft_embedding', 'logits_probability_chunk');
|
|
290
|
+
const uniformBuffer = createLogitsUniformBuffer(
|
|
291
|
+
device,
|
|
292
|
+
numTokens,
|
|
293
|
+
hiddenSize,
|
|
294
|
+
vocabSize,
|
|
295
|
+
rowStart,
|
|
296
|
+
rowCount,
|
|
297
|
+
temperature
|
|
298
|
+
);
|
|
299
|
+
try {
|
|
300
|
+
const bindGroup = await createBindGroupWithValidation(device, {
|
|
301
|
+
label: 'soft_embedding_logits_probability_chunk_bind_group',
|
|
302
|
+
layout: pipeline.getBindGroupLayout(0),
|
|
303
|
+
entries: [
|
|
304
|
+
{ binding: 0, resource: { buffer: uniformBuffer } },
|
|
305
|
+
{ binding: 1, resource: { buffer: logitsTensor.buffer } },
|
|
306
|
+
{ binding: 2, resource: { buffer: rowMaxBuffer } },
|
|
307
|
+
{ binding: 3, resource: { buffer: rowSumBuffer } },
|
|
308
|
+
{ binding: 4, resource: { buffer: probabilityBuffer } },
|
|
309
|
+
],
|
|
310
|
+
}, 'soft_embedding/logits_probability_chunk');
|
|
311
|
+
dispatchKernel(
|
|
312
|
+
null,
|
|
313
|
+
pipeline,
|
|
314
|
+
bindGroup,
|
|
315
|
+
[Math.ceil(rowCount / LOGITS_WORKGROUP_SIZE), numTokens, 1],
|
|
316
|
+
'soft_embedding_logits_probability_chunk'
|
|
317
|
+
);
|
|
318
|
+
} finally {
|
|
319
|
+
uniformBuffer.destroy();
|
|
320
|
+
}
|
|
321
|
+
}
|
|
322
|
+
|
|
323
|
+
export async function runSoftEmbeddingLogitsF16(logitsTensor, embedding, numTokens, hiddenSize, vocabSize, options = {}) {
|
|
324
|
+
validatePositiveInt(numTokens, 'numTokens');
|
|
325
|
+
validatePositiveInt(hiddenSize, 'hiddenSize');
|
|
326
|
+
validatePositiveInt(vocabSize, 'vocabSize');
|
|
327
|
+
const temperature = options.temperature ?? 1.0;
|
|
328
|
+
validatePositiveNumber(temperature, 'temperature');
|
|
329
|
+
|
|
330
|
+
if (logitsTensor?.dtype !== 'f32') {
|
|
331
|
+
throw new Error(`[SoftEmbedding] logits f16 path requires f32 logits input, got "${logitsTensor?.dtype ?? 'missing'}".`);
|
|
332
|
+
}
|
|
333
|
+
if (!isWeightBuffer(embedding)) {
|
|
334
|
+
throw new Error('[SoftEmbedding] logits f16 path requires a WeightBuffer embedding table.');
|
|
335
|
+
}
|
|
336
|
+
if (getWeightDtype(embedding) !== 'f16' || getLayout(embedding) !== 'row') {
|
|
337
|
+
throw new Error(
|
|
338
|
+
`[SoftEmbedding] logits path supports row-major f16 embeddings only; ` +
|
|
339
|
+
`got dtype=${getWeightDtype(embedding)}, layout=${getLayout(embedding)}.`
|
|
340
|
+
);
|
|
341
|
+
}
|
|
342
|
+
const shape = Array.isArray(embedding.shape) ? embedding.shape : null;
|
|
343
|
+
if (!shape || shape.length !== 2 || shape[0] !== vocabSize || shape[1] !== hiddenSize) {
|
|
344
|
+
throw new Error(
|
|
345
|
+
`[SoftEmbedding] embedding shape mismatch: expected [${vocabSize}, ${hiddenSize}], ` +
|
|
346
|
+
`got ${shape ? `[${shape.join(', ')}]` : 'missing'}.`
|
|
347
|
+
);
|
|
348
|
+
}
|
|
349
|
+
|
|
350
|
+
const device = getDevice();
|
|
351
|
+
if (!device) {
|
|
352
|
+
throw new Error('[SoftEmbedding] GPU device not available.');
|
|
353
|
+
}
|
|
354
|
+
|
|
355
|
+
const statsBytes = numTokens * Float32Array.BYTES_PER_ELEMENT;
|
|
356
|
+
const rowMaxBuffer = acquireBuffer(statsBytes, undefined, 'soft_embedding_logits_row_max');
|
|
357
|
+
const rowSumBuffer = acquireBuffer(statsBytes, undefined, 'soft_embedding_logits_row_sum');
|
|
358
|
+
let output = null;
|
|
359
|
+
let completed = false;
|
|
360
|
+
|
|
361
|
+
try {
|
|
362
|
+
await dispatchLogitsNormStats({
|
|
363
|
+
device,
|
|
364
|
+
logitsTensor,
|
|
365
|
+
rowMaxBuffer,
|
|
366
|
+
rowSumBuffer,
|
|
367
|
+
numTokens,
|
|
368
|
+
hiddenSize,
|
|
369
|
+
vocabSize,
|
|
370
|
+
temperature,
|
|
371
|
+
});
|
|
372
|
+
|
|
373
|
+
const chunkRows = resolveLogitsChunkRows(
|
|
374
|
+
device,
|
|
375
|
+
numTokens,
|
|
376
|
+
hiddenSize,
|
|
377
|
+
vocabSize,
|
|
378
|
+
options.chunkRows
|
|
379
|
+
);
|
|
380
|
+
for (let rowStart = 0; rowStart < vocabSize; rowStart += chunkRows) {
|
|
381
|
+
const rowCount = Math.min(chunkRows, vocabSize - rowStart);
|
|
382
|
+
const probabilityBytes = numTokens * rowCount * Float32Array.BYTES_PER_ELEMENT;
|
|
383
|
+
const probabilityBuffer = acquireBuffer(probabilityBytes, undefined, 'soft_embedding_logits_probability_chunk');
|
|
384
|
+
let chunkOutput = null;
|
|
385
|
+
try {
|
|
386
|
+
await dispatchProbabilityChunk({
|
|
387
|
+
device,
|
|
388
|
+
logitsTensor,
|
|
389
|
+
rowMaxBuffer,
|
|
390
|
+
rowSumBuffer,
|
|
391
|
+
probabilityBuffer,
|
|
392
|
+
numTokens,
|
|
393
|
+
hiddenSize,
|
|
394
|
+
vocabSize,
|
|
395
|
+
rowStart,
|
|
396
|
+
rowCount,
|
|
397
|
+
temperature,
|
|
398
|
+
});
|
|
399
|
+
const probabilityTensor = createTensor(
|
|
400
|
+
probabilityBuffer,
|
|
401
|
+
'f32',
|
|
402
|
+
[numTokens, rowCount],
|
|
403
|
+
'soft_embedding_logits_probability_chunk'
|
|
404
|
+
);
|
|
405
|
+
chunkOutput = await runMatmul(
|
|
406
|
+
probabilityTensor,
|
|
407
|
+
embedding,
|
|
408
|
+
numTokens,
|
|
409
|
+
hiddenSize,
|
|
410
|
+
rowCount,
|
|
411
|
+
{
|
|
412
|
+
bOffset: rowStart * hiddenSize * Uint16Array.BYTES_PER_ELEMENT,
|
|
413
|
+
transposeB: false,
|
|
414
|
+
role: 'diffusion_gemma_self_conditioning_embed',
|
|
415
|
+
outputDtype: 'f32',
|
|
416
|
+
}
|
|
417
|
+
);
|
|
418
|
+
if (!output) {
|
|
419
|
+
output = chunkOutput;
|
|
420
|
+
chunkOutput = null;
|
|
421
|
+
} else {
|
|
422
|
+
const nextOutput = await runResidualAdd(output, chunkOutput, numTokens * hiddenSize, {
|
|
423
|
+
useVec4: true,
|
|
424
|
+
});
|
|
425
|
+
releaseBuffer(output.buffer);
|
|
426
|
+
releaseBuffer(chunkOutput.buffer);
|
|
427
|
+
output = nextOutput;
|
|
428
|
+
chunkOutput = null;
|
|
429
|
+
}
|
|
430
|
+
} finally {
|
|
431
|
+
releaseBuffer(probabilityBuffer);
|
|
432
|
+
if (chunkOutput?.buffer) {
|
|
433
|
+
releaseBuffer(chunkOutput.buffer);
|
|
434
|
+
}
|
|
435
|
+
}
|
|
436
|
+
}
|
|
437
|
+
|
|
438
|
+
completed = true;
|
|
439
|
+
return createTensor(output.buffer, 'f32', [numTokens, hiddenSize], 'soft_embedding_logits_output');
|
|
440
|
+
} finally {
|
|
441
|
+
releaseBuffer(rowMaxBuffer);
|
|
442
|
+
releaseBuffer(rowSumBuffer);
|
|
443
|
+
if (!completed && output?.buffer) {
|
|
444
|
+
releaseBuffer(output.buffer);
|
|
445
|
+
}
|
|
446
|
+
}
|
|
447
|
+
}
|
|
@@ -0,0 +1,107 @@
|
|
|
1
|
+
enable f16;
|
|
2
|
+
|
|
3
|
+
override WORKGROUP_SIZE: u32 = 256u;
|
|
4
|
+
const NEG_INF: f32 = -3.402823e+38;
|
|
5
|
+
|
|
6
|
+
struct Uniforms {
|
|
7
|
+
canvas_length: u32,
|
|
8
|
+
hidden_size: u32,
|
|
9
|
+
vocab_size: u32,
|
|
10
|
+
row_start: u32,
|
|
11
|
+
row_count: u32,
|
|
12
|
+
temperature: f32,
|
|
13
|
+
_pad0: u32,
|
|
14
|
+
_pad1: u32,
|
|
15
|
+
}
|
|
16
|
+
|
|
17
|
+
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
18
|
+
@group(0) @binding(1) var<storage, read> logits: array<f32>;
|
|
19
|
+
@group(0) @binding(2) var<storage, read_write> row_max: array<f32>;
|
|
20
|
+
@group(0) @binding(3) var<storage, read_write> row_sum: array<f32>;
|
|
21
|
+
@group(0) @binding(4) var<storage, read_write> probabilities: array<f32>;
|
|
22
|
+
|
|
23
|
+
var<workgroup> shared_max: array<f32, 256>;
|
|
24
|
+
var<workgroup> shared_sum: array<f32, 256>;
|
|
25
|
+
|
|
26
|
+
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
27
|
+
fn logits_norm_stats(
|
|
28
|
+
@builtin(local_invocation_id) local_id: vec3<u32>,
|
|
29
|
+
@builtin(workgroup_id) workgroup_id: vec3<u32>
|
|
30
|
+
) {
|
|
31
|
+
let row = workgroup_id.x;
|
|
32
|
+
let thread_idx = local_id.x;
|
|
33
|
+
if (row >= u.canvas_length) {
|
|
34
|
+
return;
|
|
35
|
+
}
|
|
36
|
+
|
|
37
|
+
let base = row * u.vocab_size;
|
|
38
|
+
let temperature = max(u.temperature, 0.000001);
|
|
39
|
+
|
|
40
|
+
var local_max = NEG_INF;
|
|
41
|
+
var idx = thread_idx;
|
|
42
|
+
while (idx < u.vocab_size) {
|
|
43
|
+
local_max = max(local_max, logits[base + idx] / temperature);
|
|
44
|
+
idx = idx + WORKGROUP_SIZE;
|
|
45
|
+
}
|
|
46
|
+
|
|
47
|
+
shared_max[thread_idx] = local_max;
|
|
48
|
+
workgroupBarrier();
|
|
49
|
+
|
|
50
|
+
var stride = WORKGROUP_SIZE / 2u;
|
|
51
|
+
while (stride > 0u) {
|
|
52
|
+
if (thread_idx < stride) {
|
|
53
|
+
shared_max[thread_idx] = max(shared_max[thread_idx], shared_max[thread_idx + stride]);
|
|
54
|
+
}
|
|
55
|
+
workgroupBarrier();
|
|
56
|
+
stride = stride / 2u;
|
|
57
|
+
}
|
|
58
|
+
|
|
59
|
+
let max_value = shared_max[0];
|
|
60
|
+
var local_sum = 0.0;
|
|
61
|
+
idx = thread_idx;
|
|
62
|
+
while (idx < u.vocab_size) {
|
|
63
|
+
let value = logits[base + idx] / temperature;
|
|
64
|
+
local_sum = local_sum + exp(value - max_value);
|
|
65
|
+
idx = idx + WORKGROUP_SIZE;
|
|
66
|
+
}
|
|
67
|
+
|
|
68
|
+
shared_sum[thread_idx] = local_sum;
|
|
69
|
+
workgroupBarrier();
|
|
70
|
+
|
|
71
|
+
stride = WORKGROUP_SIZE / 2u;
|
|
72
|
+
while (stride > 0u) {
|
|
73
|
+
if (thread_idx < stride) {
|
|
74
|
+
shared_sum[thread_idx] = shared_sum[thread_idx] + shared_sum[thread_idx + stride];
|
|
75
|
+
}
|
|
76
|
+
workgroupBarrier();
|
|
77
|
+
stride = stride / 2u;
|
|
78
|
+
}
|
|
79
|
+
|
|
80
|
+
if (thread_idx == 0u) {
|
|
81
|
+
row_max[row] = max_value;
|
|
82
|
+
row_sum[row] = shared_sum[0];
|
|
83
|
+
}
|
|
84
|
+
}
|
|
85
|
+
|
|
86
|
+
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
87
|
+
fn logits_probability_chunk(@builtin(global_invocation_id) gid: vec3<u32>) {
|
|
88
|
+
let chunk_idx = gid.x;
|
|
89
|
+
let row = gid.y;
|
|
90
|
+
if (row >= u.canvas_length || chunk_idx >= u.row_count) {
|
|
91
|
+
return;
|
|
92
|
+
}
|
|
93
|
+
|
|
94
|
+
let vocab_idx = u.row_start + chunk_idx;
|
|
95
|
+
if (vocab_idx >= u.vocab_size) {
|
|
96
|
+
return;
|
|
97
|
+
}
|
|
98
|
+
|
|
99
|
+
let sum = row_sum[row];
|
|
100
|
+
let temperature = max(u.temperature, 0.000001);
|
|
101
|
+
let value = logits[row * u.vocab_size + vocab_idx] / temperature;
|
|
102
|
+
probabilities[row * u.row_count + chunk_idx] = select(
|
|
103
|
+
0.0,
|
|
104
|
+
exp(value - row_max[row]) / sum,
|
|
105
|
+
sum > 0.0
|
|
106
|
+
);
|
|
107
|
+
}
|