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 +1,116 @@
|
|
|
1
|
+
enable f16;
|
|
2
|
+
|
|
3
|
+
override WORKGROUP_SIZE: u32 = 256u;
|
|
4
|
+
override COLS_PER_WG: u32 = 8u;
|
|
5
|
+
override LANES_PER_COL: u32 = 32u;
|
|
6
|
+
override TILE_M: u32 = 4u;
|
|
7
|
+
override SCALE_DTYPE: u32 = 1u;
|
|
8
|
+
|
|
9
|
+
const MAX_WORKGROUP_SIZE: u32 = 256u;
|
|
10
|
+
const MAX_TILE_M: u32 = 4u;
|
|
11
|
+
|
|
12
|
+
struct Uniforms {
|
|
13
|
+
M: u32,
|
|
14
|
+
N: u32,
|
|
15
|
+
K: u32,
|
|
16
|
+
alpha: f32,
|
|
17
|
+
transpose_b: u32,
|
|
18
|
+
workgroups_x: u32,
|
|
19
|
+
_pad1: u32,
|
|
20
|
+
_pad2: u32,
|
|
21
|
+
}
|
|
22
|
+
|
|
23
|
+
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
24
|
+
@group(0) @binding(1) var<storage, read> A: array<f16>;
|
|
25
|
+
@group(0) @binding(2) var<storage, read> B_words: array<u32>;
|
|
26
|
+
@group(0) @binding(3) var<storage, read> S_words: array<u32>;
|
|
27
|
+
@group(0) @binding(4) var<storage, read_write> C_f16: array<f16>;
|
|
28
|
+
|
|
29
|
+
var<workgroup> partials: array<f32, MAX_WORKGROUP_SIZE * MAX_TILE_M>;
|
|
30
|
+
|
|
31
|
+
fn packed_byte(byte_index: u32) -> u32 {
|
|
32
|
+
let word = B_words[byte_index >> 2u];
|
|
33
|
+
let shift = (byte_index & 3u) * 8u;
|
|
34
|
+
return (word >> shift) & 0xffu;
|
|
35
|
+
}
|
|
36
|
+
|
|
37
|
+
fn packed_scale_half(scale_index: u32) -> u32 {
|
|
38
|
+
let word = S_words[scale_index >> 1u];
|
|
39
|
+
return select(word & 0xffffu, (word >> 16u) & 0xffffu, (scale_index & 1u) == 1u);
|
|
40
|
+
}
|
|
41
|
+
|
|
42
|
+
fn scale_value(scale_index: u32) -> f32 {
|
|
43
|
+
if (SCALE_DTYPE == 2u) {
|
|
44
|
+
return bitcast<f32>(S_words[scale_index]);
|
|
45
|
+
}
|
|
46
|
+
let half = packed_scale_half(scale_index);
|
|
47
|
+
if (SCALE_DTYPE == 1u) {
|
|
48
|
+
return bitcast<f32>(half << 16u);
|
|
49
|
+
}
|
|
50
|
+
let pair = unpack2x16float(S_words[scale_index >> 1u]);
|
|
51
|
+
return select(pair.x, pair.y, (scale_index & 1u) == 1u);
|
|
52
|
+
}
|
|
53
|
+
|
|
54
|
+
fn offset_binary_int4(byte_value: u32, high_nibble: bool) -> f32 {
|
|
55
|
+
let value = select(byte_value & 0x0fu, (byte_value >> 4u) & 0x0fu, high_nibble);
|
|
56
|
+
return f32(value) - 8.0;
|
|
57
|
+
}
|
|
58
|
+
|
|
59
|
+
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
60
|
+
fn main_batched_f16a(
|
|
61
|
+
@builtin(local_invocation_id) lid: vec3<u32>,
|
|
62
|
+
@builtin(workgroup_id) wg_id: vec3<u32>
|
|
63
|
+
) {
|
|
64
|
+
if (TILE_M > MAX_TILE_M) {
|
|
65
|
+
return;
|
|
66
|
+
}
|
|
67
|
+
|
|
68
|
+
let local_id = lid.x;
|
|
69
|
+
let col_in_wg = local_id / LANES_PER_COL;
|
|
70
|
+
let lane_in_col = local_id - col_in_wg * LANES_PER_COL;
|
|
71
|
+
let col = wg_id.x * COLS_PER_WG + col_in_wg;
|
|
72
|
+
let row_base_out = wg_id.y * TILE_M;
|
|
73
|
+
let groups_per_row = (u.K + 31u) >> 5u;
|
|
74
|
+
|
|
75
|
+
var sums: array<f32, MAX_TILE_M>;
|
|
76
|
+
for (var row_in_tile: u32 = 0u; row_in_tile < TILE_M; row_in_tile = row_in_tile + 1u) {
|
|
77
|
+
sums[row_in_tile] = 0.0;
|
|
78
|
+
}
|
|
79
|
+
|
|
80
|
+
if (col < u.N) {
|
|
81
|
+
let weight_base = col * groups_per_row * 16u;
|
|
82
|
+
let scale_base = col * groups_per_row;
|
|
83
|
+
for (var k: u32 = lane_in_col; k < u.K; k = k + LANES_PER_COL) {
|
|
84
|
+
let group = k >> 5u;
|
|
85
|
+
let lane = k & 31u;
|
|
86
|
+
let byte_value = packed_byte(weight_base + group * 16u + (lane >> 1u));
|
|
87
|
+
let weight = offset_binary_int4(byte_value, (lane & 1u) == 1u) * scale_value(scale_base + group);
|
|
88
|
+
for (var row_in_tile: u32 = 0u; row_in_tile < TILE_M; row_in_tile = row_in_tile + 1u) {
|
|
89
|
+
let row = row_base_out + row_in_tile;
|
|
90
|
+
if (row < u.M) {
|
|
91
|
+
sums[row_in_tile] = sums[row_in_tile] + f32(A[row * u.K + k]) * weight;
|
|
92
|
+
}
|
|
93
|
+
}
|
|
94
|
+
}
|
|
95
|
+
}
|
|
96
|
+
|
|
97
|
+
for (var row_in_tile: u32 = 0u; row_in_tile < TILE_M; row_in_tile = row_in_tile + 1u) {
|
|
98
|
+
partials[row_in_tile * WORKGROUP_SIZE + local_id] = sums[row_in_tile];
|
|
99
|
+
}
|
|
100
|
+
workgroupBarrier();
|
|
101
|
+
|
|
102
|
+
if (lane_in_col == 0u && col < u.N) {
|
|
103
|
+
let lane_base = col_in_wg * LANES_PER_COL;
|
|
104
|
+
for (var row_in_tile: u32 = 0u; row_in_tile < TILE_M; row_in_tile = row_in_tile + 1u) {
|
|
105
|
+
let row = row_base_out + row_in_tile;
|
|
106
|
+
if (row < u.M) {
|
|
107
|
+
var final_sum: f32 = 0.0;
|
|
108
|
+
let partial_base = row_in_tile * WORKGROUP_SIZE + lane_base;
|
|
109
|
+
for (var lane: u32 = 0u; lane < LANES_PER_COL; lane = lane + 1u) {
|
|
110
|
+
final_sum = final_sum + partials[partial_base + lane];
|
|
111
|
+
}
|
|
112
|
+
C_f16[row * u.N + col] = f16(final_sum * u.alpha);
|
|
113
|
+
}
|
|
114
|
+
}
|
|
115
|
+
}
|
|
116
|
+
}
|
|
@@ -0,0 +1,116 @@
|
|
|
1
|
+
enable f16;
|
|
2
|
+
|
|
3
|
+
override WORKGROUP_SIZE: u32 = 256u;
|
|
4
|
+
override COLS_PER_WG: u32 = 8u;
|
|
5
|
+
override LANES_PER_COL: u32 = 32u;
|
|
6
|
+
override TILE_M: u32 = 4u;
|
|
7
|
+
override SCALE_DTYPE: u32 = 1u;
|
|
8
|
+
|
|
9
|
+
const MAX_WORKGROUP_SIZE: u32 = 256u;
|
|
10
|
+
const MAX_TILE_M: u32 = 4u;
|
|
11
|
+
|
|
12
|
+
struct Uniforms {
|
|
13
|
+
M: u32,
|
|
14
|
+
N: u32,
|
|
15
|
+
K: u32,
|
|
16
|
+
alpha: f32,
|
|
17
|
+
transpose_b: u32,
|
|
18
|
+
workgroups_x: u32,
|
|
19
|
+
_pad1: u32,
|
|
20
|
+
_pad2: u32,
|
|
21
|
+
}
|
|
22
|
+
|
|
23
|
+
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
24
|
+
@group(0) @binding(1) var<storage, read> A: array<f16>;
|
|
25
|
+
@group(0) @binding(2) var<storage, read> B_words: array<u32>;
|
|
26
|
+
@group(0) @binding(3) var<storage, read> S_words: array<u32>;
|
|
27
|
+
@group(0) @binding(4) var<storage, read_write> C_f32: array<f32>;
|
|
28
|
+
|
|
29
|
+
var<workgroup> partials: array<f32, MAX_WORKGROUP_SIZE * MAX_TILE_M>;
|
|
30
|
+
|
|
31
|
+
fn packed_byte(byte_index: u32) -> u32 {
|
|
32
|
+
let word = B_words[byte_index >> 2u];
|
|
33
|
+
let shift = (byte_index & 3u) * 8u;
|
|
34
|
+
return (word >> shift) & 0xffu;
|
|
35
|
+
}
|
|
36
|
+
|
|
37
|
+
fn packed_scale_half(scale_index: u32) -> u32 {
|
|
38
|
+
let word = S_words[scale_index >> 1u];
|
|
39
|
+
return select(word & 0xffffu, (word >> 16u) & 0xffffu, (scale_index & 1u) == 1u);
|
|
40
|
+
}
|
|
41
|
+
|
|
42
|
+
fn scale_value(scale_index: u32) -> f32 {
|
|
43
|
+
if (SCALE_DTYPE == 2u) {
|
|
44
|
+
return bitcast<f32>(S_words[scale_index]);
|
|
45
|
+
}
|
|
46
|
+
let half = packed_scale_half(scale_index);
|
|
47
|
+
if (SCALE_DTYPE == 1u) {
|
|
48
|
+
return bitcast<f32>(half << 16u);
|
|
49
|
+
}
|
|
50
|
+
let pair = unpack2x16float(S_words[scale_index >> 1u]);
|
|
51
|
+
return select(pair.x, pair.y, (scale_index & 1u) == 1u);
|
|
52
|
+
}
|
|
53
|
+
|
|
54
|
+
fn offset_binary_int4(byte_value: u32, high_nibble: bool) -> f32 {
|
|
55
|
+
let value = select(byte_value & 0x0fu, (byte_value >> 4u) & 0x0fu, high_nibble);
|
|
56
|
+
return f32(value) - 8.0;
|
|
57
|
+
}
|
|
58
|
+
|
|
59
|
+
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
60
|
+
fn main_batched_f16a_f32out(
|
|
61
|
+
@builtin(local_invocation_id) lid: vec3<u32>,
|
|
62
|
+
@builtin(workgroup_id) wg_id: vec3<u32>
|
|
63
|
+
) {
|
|
64
|
+
if (TILE_M > MAX_TILE_M) {
|
|
65
|
+
return;
|
|
66
|
+
}
|
|
67
|
+
|
|
68
|
+
let local_id = lid.x;
|
|
69
|
+
let col_in_wg = local_id / LANES_PER_COL;
|
|
70
|
+
let lane_in_col = local_id - col_in_wg * LANES_PER_COL;
|
|
71
|
+
let col = wg_id.x * COLS_PER_WG + col_in_wg;
|
|
72
|
+
let row_base_out = wg_id.y * TILE_M;
|
|
73
|
+
let groups_per_row = (u.K + 31u) >> 5u;
|
|
74
|
+
|
|
75
|
+
var sums: array<f32, MAX_TILE_M>;
|
|
76
|
+
for (var row_in_tile: u32 = 0u; row_in_tile < TILE_M; row_in_tile = row_in_tile + 1u) {
|
|
77
|
+
sums[row_in_tile] = 0.0;
|
|
78
|
+
}
|
|
79
|
+
|
|
80
|
+
if (col < u.N) {
|
|
81
|
+
let weight_base = col * groups_per_row * 16u;
|
|
82
|
+
let scale_base = col * groups_per_row;
|
|
83
|
+
for (var k: u32 = lane_in_col; k < u.K; k = k + LANES_PER_COL) {
|
|
84
|
+
let group = k >> 5u;
|
|
85
|
+
let lane = k & 31u;
|
|
86
|
+
let byte_value = packed_byte(weight_base + group * 16u + (lane >> 1u));
|
|
87
|
+
let weight = offset_binary_int4(byte_value, (lane & 1u) == 1u) * scale_value(scale_base + group);
|
|
88
|
+
for (var row_in_tile: u32 = 0u; row_in_tile < TILE_M; row_in_tile = row_in_tile + 1u) {
|
|
89
|
+
let row = row_base_out + row_in_tile;
|
|
90
|
+
if (row < u.M) {
|
|
91
|
+
sums[row_in_tile] = sums[row_in_tile] + f32(A[row * u.K + k]) * weight;
|
|
92
|
+
}
|
|
93
|
+
}
|
|
94
|
+
}
|
|
95
|
+
}
|
|
96
|
+
|
|
97
|
+
for (var row_in_tile: u32 = 0u; row_in_tile < TILE_M; row_in_tile = row_in_tile + 1u) {
|
|
98
|
+
partials[row_in_tile * WORKGROUP_SIZE + local_id] = sums[row_in_tile];
|
|
99
|
+
}
|
|
100
|
+
workgroupBarrier();
|
|
101
|
+
|
|
102
|
+
if (lane_in_col == 0u && col < u.N) {
|
|
103
|
+
let lane_base = col_in_wg * LANES_PER_COL;
|
|
104
|
+
for (var row_in_tile: u32 = 0u; row_in_tile < TILE_M; row_in_tile = row_in_tile + 1u) {
|
|
105
|
+
let row = row_base_out + row_in_tile;
|
|
106
|
+
if (row < u.M) {
|
|
107
|
+
var final_sum: f32 = 0.0;
|
|
108
|
+
let partial_base = row_in_tile * WORKGROUP_SIZE + lane_base;
|
|
109
|
+
for (var lane: u32 = 0u; lane < LANES_PER_COL; lane = lane + 1u) {
|
|
110
|
+
final_sum = final_sum + partials[partial_base + lane];
|
|
111
|
+
}
|
|
112
|
+
C_f32[row * u.N + col] = final_sum * u.alpha;
|
|
113
|
+
}
|
|
114
|
+
}
|
|
115
|
+
}
|
|
116
|
+
}
|
|
@@ -0,0 +1,95 @@
|
|
|
1
|
+
enable f16;
|
|
2
|
+
|
|
3
|
+
override WORKGROUP_SIZE: u32 = 256u;
|
|
4
|
+
override COLS_PER_WG: u32 = 8u;
|
|
5
|
+
override MULTICOL_COLS_PER_WG: u32 = 8u;
|
|
6
|
+
override MULTICOL_THREADS_PER_COL: u32 = 32u;
|
|
7
|
+
override SCALE_DTYPE: u32 = 1u;
|
|
8
|
+
|
|
9
|
+
const MAX_WORKGROUP_SIZE: u32 = 256u;
|
|
10
|
+
|
|
11
|
+
struct Uniforms {
|
|
12
|
+
M: u32,
|
|
13
|
+
N: u32,
|
|
14
|
+
K: u32,
|
|
15
|
+
alpha: f32,
|
|
16
|
+
transpose_b: u32,
|
|
17
|
+
workgroups_x: u32,
|
|
18
|
+
_pad1: u32,
|
|
19
|
+
_pad2: u32,
|
|
20
|
+
}
|
|
21
|
+
|
|
22
|
+
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
23
|
+
@group(0) @binding(1) var<storage, read> A: array<f16>;
|
|
24
|
+
@group(0) @binding(2) var<storage, read> B_words: array<u32>;
|
|
25
|
+
@group(0) @binding(3) var<storage, read> S_words: array<u32>;
|
|
26
|
+
@group(0) @binding(4) var<storage, read_write> C_f16: array<f16>;
|
|
27
|
+
|
|
28
|
+
var<workgroup> partials: array<f32, MAX_WORKGROUP_SIZE>;
|
|
29
|
+
|
|
30
|
+
fn packed_byte(byte_index: u32) -> u32 {
|
|
31
|
+
let word = B_words[byte_index >> 2u];
|
|
32
|
+
let shift = (byte_index & 3u) * 8u;
|
|
33
|
+
return (word >> shift) & 0xffu;
|
|
34
|
+
}
|
|
35
|
+
|
|
36
|
+
fn packed_scale_half(scale_index: u32) -> u32 {
|
|
37
|
+
let word = S_words[scale_index >> 1u];
|
|
38
|
+
return select(word & 0xffffu, (word >> 16u) & 0xffffu, (scale_index & 1u) == 1u);
|
|
39
|
+
}
|
|
40
|
+
|
|
41
|
+
fn scale_value(scale_index: u32) -> f32 {
|
|
42
|
+
if (SCALE_DTYPE == 2u) {
|
|
43
|
+
return bitcast<f32>(S_words[scale_index]);
|
|
44
|
+
}
|
|
45
|
+
let half = packed_scale_half(scale_index);
|
|
46
|
+
if (SCALE_DTYPE == 1u) {
|
|
47
|
+
return bitcast<f32>(half << 16u);
|
|
48
|
+
}
|
|
49
|
+
let pair = unpack2x16float(S_words[scale_index >> 1u]);
|
|
50
|
+
return select(pair.x, pair.y, (scale_index & 1u) == 1u);
|
|
51
|
+
}
|
|
52
|
+
|
|
53
|
+
fn offset_binary_int4(byte_value: u32, high_nibble: bool) -> f32 {
|
|
54
|
+
let value = select(byte_value & 0x0fu, (byte_value >> 4u) & 0x0fu, high_nibble);
|
|
55
|
+
return f32(value) - 8.0;
|
|
56
|
+
}
|
|
57
|
+
|
|
58
|
+
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
59
|
+
fn main_multicol_f16a(
|
|
60
|
+
@builtin(local_invocation_id) lid: vec3<u32>,
|
|
61
|
+
@builtin(workgroup_id) wg_id: vec3<u32>
|
|
62
|
+
) {
|
|
63
|
+
let local_id = lid.x;
|
|
64
|
+
let lanes_per_col = WORKGROUP_SIZE / MULTICOL_COLS_PER_WG;
|
|
65
|
+
let col_in_wg = local_id / lanes_per_col;
|
|
66
|
+
let lane_in_col = local_id - col_in_wg * lanes_per_col;
|
|
67
|
+
let row = wg_id.y;
|
|
68
|
+
let col = wg_id.x * MULTICOL_COLS_PER_WG + col_in_wg;
|
|
69
|
+
let groups_per_row = (u.K + 31u) >> 5u;
|
|
70
|
+
|
|
71
|
+
var sum: f32 = 0.0;
|
|
72
|
+
if (row < u.M && col < u.N) {
|
|
73
|
+
let row_base = col * groups_per_row * 16u;
|
|
74
|
+
let scale_base = col * groups_per_row;
|
|
75
|
+
for (var k: u32 = lane_in_col; k < u.K; k = k + lanes_per_col) {
|
|
76
|
+
let group = k >> 5u;
|
|
77
|
+
let lane = k & 31u;
|
|
78
|
+
let byte_value = packed_byte(row_base + group * 16u + (lane >> 1u));
|
|
79
|
+
let weight = offset_binary_int4(byte_value, (lane & 1u) == 1u) * scale_value(scale_base + group);
|
|
80
|
+
sum = sum + f32(A[row * u.K + k]) * weight;
|
|
81
|
+
}
|
|
82
|
+
}
|
|
83
|
+
|
|
84
|
+
partials[local_id] = sum;
|
|
85
|
+
workgroupBarrier();
|
|
86
|
+
|
|
87
|
+
if (lane_in_col == 0u && row < u.M && col < u.N) {
|
|
88
|
+
var final_sum: f32 = 0.0;
|
|
89
|
+
let base = col_in_wg * lanes_per_col;
|
|
90
|
+
for (var lane: u32 = 0u; lane < lanes_per_col; lane = lane + 1u) {
|
|
91
|
+
final_sum = final_sum + partials[base + lane];
|
|
92
|
+
}
|
|
93
|
+
C_f16[row * u.N + col] = f16(final_sum * u.alpha);
|
|
94
|
+
}
|
|
95
|
+
}
|
|
@@ -0,0 +1,95 @@
|
|
|
1
|
+
enable f16;
|
|
2
|
+
|
|
3
|
+
override WORKGROUP_SIZE: u32 = 256u;
|
|
4
|
+
override COLS_PER_WG: u32 = 8u;
|
|
5
|
+
override MULTICOL_COLS_PER_WG: u32 = 8u;
|
|
6
|
+
override MULTICOL_THREADS_PER_COL: u32 = 32u;
|
|
7
|
+
override SCALE_DTYPE: u32 = 1u;
|
|
8
|
+
|
|
9
|
+
const MAX_WORKGROUP_SIZE: u32 = 256u;
|
|
10
|
+
|
|
11
|
+
struct Uniforms {
|
|
12
|
+
M: u32,
|
|
13
|
+
N: u32,
|
|
14
|
+
K: u32,
|
|
15
|
+
alpha: f32,
|
|
16
|
+
transpose_b: u32,
|
|
17
|
+
workgroups_x: u32,
|
|
18
|
+
_pad1: u32,
|
|
19
|
+
_pad2: u32,
|
|
20
|
+
}
|
|
21
|
+
|
|
22
|
+
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
23
|
+
@group(0) @binding(1) var<storage, read> A: array<f16>;
|
|
24
|
+
@group(0) @binding(2) var<storage, read> B_words: array<u32>;
|
|
25
|
+
@group(0) @binding(3) var<storage, read> S_words: array<u32>;
|
|
26
|
+
@group(0) @binding(4) var<storage, read_write> C_f32: array<f32>;
|
|
27
|
+
|
|
28
|
+
var<workgroup> partials: array<f32, MAX_WORKGROUP_SIZE>;
|
|
29
|
+
|
|
30
|
+
fn packed_byte(byte_index: u32) -> u32 {
|
|
31
|
+
let word = B_words[byte_index >> 2u];
|
|
32
|
+
let shift = (byte_index & 3u) * 8u;
|
|
33
|
+
return (word >> shift) & 0xffu;
|
|
34
|
+
}
|
|
35
|
+
|
|
36
|
+
fn packed_scale_half(scale_index: u32) -> u32 {
|
|
37
|
+
let word = S_words[scale_index >> 1u];
|
|
38
|
+
return select(word & 0xffffu, (word >> 16u) & 0xffffu, (scale_index & 1u) == 1u);
|
|
39
|
+
}
|
|
40
|
+
|
|
41
|
+
fn scale_value(scale_index: u32) -> f32 {
|
|
42
|
+
if (SCALE_DTYPE == 2u) {
|
|
43
|
+
return bitcast<f32>(S_words[scale_index]);
|
|
44
|
+
}
|
|
45
|
+
let half = packed_scale_half(scale_index);
|
|
46
|
+
if (SCALE_DTYPE == 1u) {
|
|
47
|
+
return bitcast<f32>(half << 16u);
|
|
48
|
+
}
|
|
49
|
+
let pair = unpack2x16float(S_words[scale_index >> 1u]);
|
|
50
|
+
return select(pair.x, pair.y, (scale_index & 1u) == 1u);
|
|
51
|
+
}
|
|
52
|
+
|
|
53
|
+
fn offset_binary_int4(byte_value: u32, high_nibble: bool) -> f32 {
|
|
54
|
+
let value = select(byte_value & 0x0fu, (byte_value >> 4u) & 0x0fu, high_nibble);
|
|
55
|
+
return f32(value) - 8.0;
|
|
56
|
+
}
|
|
57
|
+
|
|
58
|
+
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
59
|
+
fn main_multicol_f16a_f32out(
|
|
60
|
+
@builtin(local_invocation_id) lid: vec3<u32>,
|
|
61
|
+
@builtin(workgroup_id) wg_id: vec3<u32>
|
|
62
|
+
) {
|
|
63
|
+
let local_id = lid.x;
|
|
64
|
+
let lanes_per_col = WORKGROUP_SIZE / MULTICOL_COLS_PER_WG;
|
|
65
|
+
let col_in_wg = local_id / lanes_per_col;
|
|
66
|
+
let lane_in_col = local_id - col_in_wg * lanes_per_col;
|
|
67
|
+
let row = wg_id.y;
|
|
68
|
+
let col = wg_id.x * MULTICOL_COLS_PER_WG + col_in_wg;
|
|
69
|
+
let groups_per_row = (u.K + 31u) >> 5u;
|
|
70
|
+
|
|
71
|
+
var sum: f32 = 0.0;
|
|
72
|
+
if (row < u.M && col < u.N) {
|
|
73
|
+
let row_base = col * groups_per_row * 16u;
|
|
74
|
+
let scale_base = col * groups_per_row;
|
|
75
|
+
for (var k: u32 = lane_in_col; k < u.K; k = k + lanes_per_col) {
|
|
76
|
+
let group = k >> 5u;
|
|
77
|
+
let lane = k & 31u;
|
|
78
|
+
let byte_value = packed_byte(row_base + group * 16u + (lane >> 1u));
|
|
79
|
+
let weight = offset_binary_int4(byte_value, (lane & 1u) == 1u) * scale_value(scale_base + group);
|
|
80
|
+
sum = sum + f32(A[row * u.K + k]) * weight;
|
|
81
|
+
}
|
|
82
|
+
}
|
|
83
|
+
|
|
84
|
+
partials[local_id] = sum;
|
|
85
|
+
workgroupBarrier();
|
|
86
|
+
|
|
87
|
+
if (lane_in_col == 0u && row < u.M && col < u.N) {
|
|
88
|
+
var final_sum: f32 = 0.0;
|
|
89
|
+
let base = col_in_wg * lanes_per_col;
|
|
90
|
+
for (var lane: u32 = 0u; lane < lanes_per_col; lane = lane + 1u) {
|
|
91
|
+
final_sum = final_sum + partials[base + lane];
|
|
92
|
+
}
|
|
93
|
+
C_f32[row * u.N + col] = final_sum * u.alpha;
|
|
94
|
+
}
|
|
95
|
+
}
|
|
@@ -7,11 +7,12 @@
|
|
|
7
7
|
import type { CommandRecorder } from '../command-recorder.js';
|
|
8
8
|
import type { Tensor, TensorDtype } from '../tensor.js';
|
|
9
9
|
import type { OutputBufferOptions } from './types.js';
|
|
10
|
+
import type { SplitWeightBuffer, WeightDtype, WeightStorageEncoding } from '../weight-buffer.js';
|
|
10
11
|
|
|
11
12
|
/** Gather kernel options */
|
|
12
13
|
export interface GatherOptions extends OutputBufferOptions {
|
|
13
14
|
useVec4?: boolean;
|
|
14
|
-
embeddingDtype?: 'f16' | 'f32'
|
|
15
|
+
embeddingDtype?: Extract<WeightDtype, 'f16' | 'f32' | 'litert_int4'>;
|
|
15
16
|
/**
|
|
16
17
|
* Output dtype. When 'f16', converts F32 embeddings to F16 output.
|
|
17
18
|
* Default: 'f32'
|
|
@@ -33,6 +34,10 @@ export interface GatherOptions extends OutputBufferOptions {
|
|
|
33
34
|
indirectBuffer?: GPUBuffer | null;
|
|
34
35
|
/** Byte offset into indirect dispatch buffer (default: 0). */
|
|
35
36
|
indirectOffset?: number;
|
|
37
|
+
/** Required when embeddingDtype is 'litert_int4'. */
|
|
38
|
+
storageEncoding?: WeightStorageEncoding | null;
|
|
39
|
+
/** Optional authored split gather section count, usually stamped from the manifest-selected embedding kernel. */
|
|
40
|
+
splitGatherSectionCount?: number | null;
|
|
36
41
|
}
|
|
37
42
|
|
|
38
43
|
/**
|
|
@@ -59,3 +64,81 @@ export declare function recordGather(
|
|
|
59
64
|
vocabSize: number,
|
|
60
65
|
options?: GatherOptions
|
|
61
66
|
): Promise<Tensor>;
|
|
67
|
+
|
|
68
|
+
/**
|
|
69
|
+
* Run gather/embedding lookup against a row-split F16 embedding table.
|
|
70
|
+
*/
|
|
71
|
+
export declare function runGatherSplit4(
|
|
72
|
+
indices: GPUBuffer,
|
|
73
|
+
splitEmbedding: SplitWeightBuffer,
|
|
74
|
+
numTokens: number,
|
|
75
|
+
hiddenSize: number,
|
|
76
|
+
vocabSize: number,
|
|
77
|
+
options?: GatherOptions
|
|
78
|
+
): Promise<Tensor>;
|
|
79
|
+
|
|
80
|
+
/**
|
|
81
|
+
* Record gather against a row-split F16 embedding table.
|
|
82
|
+
*/
|
|
83
|
+
export declare function recordGatherSplit4(
|
|
84
|
+
recorder: CommandRecorder,
|
|
85
|
+
indices: GPUBuffer,
|
|
86
|
+
splitEmbedding: SplitWeightBuffer,
|
|
87
|
+
numTokens: number,
|
|
88
|
+
hiddenSize: number,
|
|
89
|
+
vocabSize: number,
|
|
90
|
+
options?: GatherOptions
|
|
91
|
+
): Promise<Tensor>;
|
|
92
|
+
|
|
93
|
+
/**
|
|
94
|
+
* Run gather/embedding lookup against a row-split F16 embedding table with up
|
|
95
|
+
* to eight sections.
|
|
96
|
+
*/
|
|
97
|
+
export declare function runGatherSplit8(
|
|
98
|
+
indices: GPUBuffer,
|
|
99
|
+
splitEmbedding: SplitWeightBuffer,
|
|
100
|
+
numTokens: number,
|
|
101
|
+
hiddenSize: number,
|
|
102
|
+
vocabSize: number,
|
|
103
|
+
options?: GatherOptions
|
|
104
|
+
): Promise<Tensor>;
|
|
105
|
+
|
|
106
|
+
/**
|
|
107
|
+
* Record gather against a row-split F16 embedding table with up to eight sections.
|
|
108
|
+
*/
|
|
109
|
+
export declare function recordGatherSplit8(
|
|
110
|
+
recorder: CommandRecorder,
|
|
111
|
+
indices: GPUBuffer,
|
|
112
|
+
splitEmbedding: SplitWeightBuffer,
|
|
113
|
+
numTokens: number,
|
|
114
|
+
hiddenSize: number,
|
|
115
|
+
vocabSize: number,
|
|
116
|
+
options?: GatherOptions
|
|
117
|
+
): Promise<Tensor>;
|
|
118
|
+
|
|
119
|
+
/**
|
|
120
|
+
* Run gather/embedding lookup against a row-split F16 embedding table, honoring
|
|
121
|
+
* the manifest-stamped split variant when the split buffer provides one.
|
|
122
|
+
*/
|
|
123
|
+
export declare function runGatherSplit(
|
|
124
|
+
indices: GPUBuffer,
|
|
125
|
+
splitEmbedding: SplitWeightBuffer,
|
|
126
|
+
numTokens: number,
|
|
127
|
+
hiddenSize: number,
|
|
128
|
+
vocabSize: number,
|
|
129
|
+
options?: GatherOptions
|
|
130
|
+
): Promise<Tensor>;
|
|
131
|
+
|
|
132
|
+
/**
|
|
133
|
+
* Record gather against a row-split F16 embedding table, honoring the
|
|
134
|
+
* manifest-stamped split variant when the split buffer provides one.
|
|
135
|
+
*/
|
|
136
|
+
export declare function recordGatherSplit(
|
|
137
|
+
recorder: CommandRecorder,
|
|
138
|
+
indices: GPUBuffer,
|
|
139
|
+
splitEmbedding: SplitWeightBuffer,
|
|
140
|
+
numTokens: number,
|
|
141
|
+
hiddenSize: number,
|
|
142
|
+
vocabSize: number,
|
|
143
|
+
options?: GatherOptions
|
|
144
|
+
): Promise<Tensor>;
|