doppler-gpu 0.4.2 → 0.4.3
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 +182 -0
- package/README.md +19 -15
- package/package.json +77 -29
- package/src/cli/cli-model-resolution.js +275 -0
- package/src/cli/cli-output.js +453 -0
- package/src/cli/doppler-cli.js +817 -709
- package/src/cli/doppler-quickstart.d.ts +7 -23
- package/src/cli/doppler-quickstart.js +4 -2
- package/src/cli/doppler-serve.js +1 -1
- package/src/cli/runtime-profiles.d.ts +40 -0
- package/src/cli/runtime-profiles.js +225 -0
- package/src/client/doppler-api.js +1 -6
- package/src/client/doppler-registry.d.ts +12 -0
- package/src/client/doppler-registry.js +27 -0
- package/src/client/doppler-registry.json +64 -6
- package/src/client/failure-taxonomy.js +139 -0
- package/src/client/fault-injection.js +30 -0
- package/src/client/provider.d.ts +151 -0
- package/src/client/provider.js +374 -0
- package/src/client/receipt.js +156 -0
- package/src/client/runtime/index.js +56 -4
- package/src/client/runtime/model-manager.d.ts +0 -14
- package/src/client/runtime/model-manager.js +2 -42
- package/src/client/runtime/model-session.js +33 -0
- package/src/client/runtime/model-source.d.ts +13 -1
- package/src/client/runtime/model-source.js +128 -2
- package/src/client/runtime/node-quickstart-cache.d.ts +22 -0
- package/src/client/runtime/node-quickstart-cache.js +353 -0
- package/src/client/runtime/source-runtime.js +1 -10
- package/src/client/runtime/types.d.ts +0 -60
- package/src/client/wrap-pipeline-handle.d.ts +20 -0
- package/src/client/wrap-pipeline-handle.js +84 -0
- package/src/config/README.md +3 -1
- package/src/config/conversion/README.md +23 -157
- package/src/config/conversion/embeddinggemma/google-embeddinggemma-300m-q4k-ehf16-af32.json +19 -4
- package/src/config/conversion/gemma3/gemma-3-1b-it-f16-af32.json +16 -9
- package/src/config/conversion/gemma3/gemma-3-1b-it-q4k-ehf16-af32.json +24 -9
- package/src/config/conversion/gemma3/gemma-3-270m-it-q4k-ehf16-af32.json +24 -9
- package/src/config/conversion/gemma3/translategemma-4b-1b-enes-q4k-ehf16-af32.json +16 -9
- package/src/config/conversion/gemma3/translategemma-4b-it-q4k-ehf16-af32.json +24 -9
- package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af16.json +719 -0
- package/src/config/conversion/gemma4/gemma-4-31b-it-text-q4k-ehf16-af32.json +715 -0
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af16-int4ple.json +589 -0
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32-int4ple.json +566 -0
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af32.json +367 -61
- package/src/config/conversion/gemma4/gemma-4-moe-q4k-ehf16-af32.json +212 -55
- package/src/config/conversion/gpt-oss-20b-f16-xmxfp4.json +216 -57
- package/src/config/conversion/janus/janus-pro-1b-text-q4k-ehaf16.json +229 -57
- package/src/config/conversion/lfm2/lfm2.5-1.2b-instruct-q4k-ehf16-af32.json +17 -10
- package/src/config/conversion/qwen3/qwen-3-5-0-8b-q4k-ehaf16.json +75 -33
- package/src/config/conversion/qwen3/qwen-3-5-2b-q4k-ehaf16.json +91 -59
- package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-eaf16.json +517 -0
- package/src/config/conversion/qwen3/qwen-3-6-27b-q4k-ehaf16.json +511 -0
- package/src/config/execution-contract-check.js +6 -5
- package/src/config/kernel-path-loader.d.ts +8 -2
- package/src/config/kernels/kernel-ref-digests.js +125 -122
- package/src/config/kernels/registry.d.ts +0 -35
- package/src/config/kernels/registry.js +5 -9
- package/src/config/kernels/registry.json +524 -350
- package/src/config/merge-helpers.d.ts +1 -1
- package/src/config/merge-helpers.js +1 -1
- package/src/config/merge.js +35 -7
- package/src/config/param-categories.d.ts +0 -9
- package/src/config/param-categories.js +1 -5
- package/src/config/param-validator.d.ts +2 -0
- package/src/config/param-validator.js +16 -0
- package/src/config/platforms/loader.d.ts +0 -27
- package/src/config/platforms/loader.js +0 -30
- package/src/config/required-inference-fields-contract-check.js +1 -7
- package/src/config/runtime/README.md +28 -0
- package/src/config/runtime/experiments/bench/gemma4-bench-q4k.json +1 -1
- package/src/config/runtime/experiments/bench/qwen3-bench-q4k-batched.json +2 -2
- package/src/config/runtime/experiments/debug/gemma4-e2b-browser-diagnose-q4k.json +85 -0
- package/src/config/runtime/platform/metal-apple-q4k.json +19 -3
- package/src/config/runtime/profiles/gemma4-31b-f16-activations-probe.json +65 -0
- package/src/config/runtime/profiles/gemma4-e2b-baseline-probe.json +41 -0
- package/src/config/runtime/profiles/gemma4-e2b-fused-rmsnorm-probe.json +44 -0
- package/src/config/runtime/profiles/gemma4-e2b-fused-rmsnorm-test.json +19 -0
- package/src/config/runtime/profiles/gemma4-e2b-int4ple-baseline-probe.json +31 -0
- package/src/config/runtime/profiles/gemma4-e2b-int4ple-f16a-probe.json +34 -0
- package/src/config/runtime/profiles/gemma4-e2b-int4ple-reference-export.json +64 -0
- package/src/config/runtime/profiles/gemma4-e2b-int4ple-retain-q4k-probe.json +47 -0
- package/src/config/runtime/profiles/gemma4-e2b-prefill-profile.json +35 -0
- package/src/config/runtime/profiles/gemma4-e2b-throughput-probe.json +16 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-prefill-profile.json +35 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-throughput-probe.json +16 -0
- package/src/config/runtime/profiles/qwen3-6-27b-f16-activations-probe.json +62 -0
- package/src/config/schema/bridge.schema.d.ts +16 -0
- package/src/config/schema/browser-suite-metrics.schema.d.ts +1 -0
- package/src/config/schema/browser-suite-metrics.schema.js +2 -0
- package/src/config/schema/converter.schema.d.ts +5 -0
- package/src/config/schema/converter.schema.js +3 -0
- package/src/config/schema/debug.schema.d.ts +22 -1
- package/src/config/schema/debug.schema.js +1 -0
- package/src/config/schema/distill-training.schema.d.ts +0 -1
- package/src/config/schema/distill-training.schema.js +1 -1
- package/src/config/schema/doppler.schema.d.ts +12 -2
- package/src/config/schema/doppler.schema.js +108 -8
- package/src/config/schema/emulation.schema.d.ts +3 -0
- package/src/config/schema/emulation.schema.js +16 -16
- package/src/config/schema/execution-v1.schema.d.ts +12 -1
- package/src/config/schema/execution-v1.schema.js +7 -1
- package/src/config/schema/index.d.ts +52 -1
- package/src/config/schema/index.js +14 -0
- package/src/config/schema/inference.schema.d.ts +3 -3
- package/src/config/schema/kernel-path.schema.d.ts +6 -27
- package/src/config/schema/kernel-thresholds.schema.d.ts +0 -2
- package/src/config/schema/kernel-thresholds.schema.js +3 -7
- package/src/config/schema/kvcache.schema.js +0 -9
- package/src/config/schema/loading.schema.d.ts +12 -0
- package/src/config/schema/loading.schema.js +4 -0
- package/src/config/schema/manifest.schema.d.ts +68 -1
- package/src/config/schema/manifest.schema.js +48 -2
- package/src/config/schema/moe.schema.js +0 -8
- package/src/config/schema/program-bundle.schema.d.ts +233 -0
- package/src/config/schema/program-bundle.schema.js +393 -0
- package/src/config/schema/ul-training.schema.d.ts +0 -1
- package/src/config/schema/ul-training.schema.js +1 -1
- package/src/config/source-packages/litert/gemma-4-e2b-it.json +36 -15
- package/src/config/supported-operations.d.ts +14 -0
- package/src/config/supported-operations.js +34 -0
- package/src/config/supported-operations.json +178 -0
- package/src/config/training-defaults.js +1 -7
- package/src/config/transforms/capability-transform-resolver.js +3 -1
- package/src/config/transforms/execution-graph-transforms.d.ts +55 -0
- package/src/config/transforms/execution-graph-transforms.js +693 -29
- package/src/converter/conversion-plan.js +13 -10
- package/src/converter/core.d.ts +57 -0
- package/src/converter/core.js +582 -13
- package/src/converter/parsers/diffusion.js +1 -33
- package/src/converter/quantization-info.d.ts +5 -4
- package/src/converter/quantization-info.js +39 -3
- package/src/converter/quantizer.d.ts +19 -0
- package/src/converter/quantizer.js +76 -0
- package/src/debug/config.d.ts +0 -4
- package/src/debug/config.js +0 -17
- package/src/debug/index.d.ts +28 -30
- package/src/debug/index.js +0 -18
- package/src/debug/stats.d.ts +0 -8
- package/src/debug/stats.js +3 -3
- package/src/debug/tensor.d.ts +9 -2
- package/src/debug/tensor.js +11 -4
- package/src/debug/trace.js +2 -15
- package/src/errors/doppler-error.js +20 -0
- package/src/experimental/browser/quantization.d.ts +5 -0
- package/src/experimental/distribution/shard-delivery.js +179 -23
- package/src/experimental/training/artifacts.js +1 -14
- package/src/experimental/training/checkpoint.js +1 -14
- package/src/experimental/training/datasets/token-batch.d.ts +1 -1
- package/src/experimental/training/datasets/token-batch.js +2 -2
- package/src/experimental/training/distillation/dataset.js +2 -15
- package/src/experimental/training/distillation/eval.js +3 -3
- package/src/experimental/training/distillation/runtime.js +3 -16
- package/src/experimental/training/distillation/student-fixture.js +6 -6
- package/src/experimental/training/lora-pipeline.js +1 -14
- package/src/experimental/training/objectives/base.d.ts +2 -2
- package/src/experimental/training/objectives/cross_entropy.d.ts +2 -2
- package/src/experimental/training/objectives/distill_kd.d.ts +2 -2
- package/src/experimental/training/objectives/distill_kd.js +2 -2
- package/src/experimental/training/objectives/distill_triplet.d.ts +2 -2
- package/src/experimental/training/objectives/distill_triplet.js +2 -2
- package/src/experimental/training/objectives/ul_stage1_joint.d.ts +2 -2
- package/src/experimental/training/objectives/ul_stage2_base.d.ts +2 -2
- package/src/experimental/training/objectives/ul_stage2_base.js +1 -1
- package/src/experimental/training/operator-artifacts.js +1 -14
- package/src/experimental/training/runner.js +1 -14
- package/src/experimental/training/workload-packs/registry.json +6 -6
- package/src/experimental/training/workloads.js +1 -1
- package/src/formats/gguf/types.d.ts +0 -63
- package/src/formats/gguf/types.js +58 -38
- package/src/formats/litert/types.d.ts +0 -13
- package/src/formats/litert/types.js +2 -2
- package/src/formats/rdrr/distributed/index.d.ts +34 -0
- package/src/formats/rdrr/distributed/index.js +3 -0
- package/src/formats/rdrr/distributed/parsing.d.ts +6 -0
- package/src/formats/rdrr/distributed/parsing.js +23 -0
- package/src/formats/rdrr/distributed/validation.d.ts +23 -0
- package/src/formats/rdrr/distributed/validation.js +131 -0
- package/src/formats/rdrr/groups.d.ts +1 -13
- package/src/formats/rdrr/groups.js +0 -34
- package/src/formats/rdrr/index.d.ts +9 -0
- package/src/formats/rdrr/index.js +9 -0
- package/src/formats/rdrr/integrity.d.ts +51 -0
- package/src/formats/rdrr/integrity.js +200 -0
- package/src/formats/rdrr/lowerings.d.ts +28 -0
- package/src/formats/rdrr/lowerings.js +112 -0
- package/src/formats/rdrr/manifest.d.ts +0 -14
- package/src/formats/rdrr/manifest.js +5 -58
- package/src/formats/rdrr/merkle.d.ts +24 -0
- package/src/formats/rdrr/merkle.js +118 -0
- package/src/formats/rdrr/parsing.js +4 -0
- package/src/formats/rdrr/slice-compiler.d.ts +29 -0
- package/src/formats/rdrr/slice-compiler.js +188 -0
- package/src/formats/rdrr/storage-descriptor.d.ts +36 -0
- package/src/formats/rdrr/storage-descriptor.js +186 -0
- package/src/formats/rdrr/tensor-config-validator.d.ts +0 -9
- package/src/formats/rdrr/tensor-config-validator.js +0 -30
- package/src/formats/rdrr/types.d.ts +107 -0
- package/src/formats/rdrr/validation.js +245 -0
- package/src/formats/tflite/types.d.ts +0 -31
- package/src/formats/tflite/types.js +5 -5
- package/src/gpu/command-recorder.d.ts +3 -21
- package/src/gpu/device.d.ts +17 -3
- package/src/gpu/device.js +20 -4
- package/src/gpu/kernel-runtime.d.ts +8 -15
- package/src/gpu/kernel-selection-log.d.ts +0 -8
- package/src/gpu/kernel-selection-log.js +0 -15
- package/src/gpu/kernel-tuner/cache.js +0 -3
- package/src/gpu/kernels/README.md +50 -18
- package/src/gpu/kernels/activation_static_qdq.wgsl +3 -1
- package/src/gpu/kernels/attention.d.ts +28 -19
- package/src/gpu/kernels/attention.js +443 -9
- package/src/gpu/kernels/attention_decode_contiguous_turboquant_f16kv.wgsl +0 -1
- package/src/gpu/kernels/attention_decode_contiguous_turboquant_prod_f16kv.wgsl +0 -1
- package/src/gpu/kernels/attention_decode_f16.wgsl +1 -1
- package/src/gpu/kernels/attention_decode_f16kv.wgsl +1 -1
- package/src/gpu/kernels/attention_decode_online_f16.wgsl +32 -29
- package/src/gpu/kernels/attention_decode_online_f16kv.wgsl +1 -1
- package/src/gpu/kernels/attention_decode_paged_f16kv.wgsl +1 -1
- package/src/gpu/kernels/attention_decode_tiered_f16kv.wgsl +1 -1
- package/src/gpu/kernels/attention_decode_tiered_turboquant_f16kv.wgsl +0 -1
- package/src/gpu/kernels/attention_decode_tiered_turboquant_prod_f16kv.wgsl +0 -1
- package/src/gpu/kernels/attention_f16.wgsl +88 -30
- package/src/gpu/kernels/attention_f16kv.wgsl +45 -15
- package/src/gpu/kernels/attention_head512_f16.wgsl +270 -0
- package/src/gpu/kernels/attention_head512_f16kv.wgsl +271 -0
- package/src/gpu/kernels/attention_prefill_flash_head256_f16kv.wgsl +306 -0
- package/src/gpu/kernels/attention_prefill_flash_reduce.wgsl +101 -0
- package/src/gpu/kernels/attention_small_f16.wgsl +22 -20
- package/src/gpu/kernels/attention_small_f16kv.wgsl +1 -1
- package/src/gpu/kernels/attention_streaming_f16.wgsl +39 -16
- package/src/gpu/kernels/attention_streaming_f16kv.wgsl +22 -3
- package/src/gpu/kernels/backward/index.d.ts +6 -0
- package/src/gpu/kernels/backward/layernorm_backward.wgsl +0 -1
- package/src/gpu/kernels/backward/utils.d.ts +35 -0
- package/src/gpu/kernels/bf16_to_f32.wgsl +0 -16
- package/src/gpu/kernels/bias_add_f16.wgsl +1 -1
- package/src/gpu/kernels/cast_f32_to_f16.wgsl +0 -1
- package/src/gpu/kernels/codegen/patches/attention_f16.from.attention.diff +19 -215
- package/src/gpu/kernels/codegen/patches/attention_f16kv.from.attention.diff +8 -82
- package/src/gpu/kernels/codegen/patches/attention_small_f16.from.attention_small.diff +92 -59
- package/src/gpu/kernels/codegen/patches/attention_streaming_f16.from.attention_streaming.diff +2 -85
- package/src/gpu/kernels/codegen/patches/gelu_f16.from.gelu.diff +27 -24
- package/src/gpu/kernels/codegen/patches/residual_f16.from.residual.diff +1 -1
- package/src/gpu/kernels/codegen/patches/rmsnorm_f16.from.rmsnorm.diff +249 -179
- package/src/gpu/kernels/codegen/patches/rope_f16.from.rope.diff +133 -235
- package/src/gpu/kernels/codegen/patches/sample_f16.from.sample.diff +119 -90
- package/src/gpu/kernels/codegen/wgsl-variants.js +142 -8
- package/src/gpu/kernels/constants.d.ts +9 -48
- package/src/gpu/kernels/constants.js +0 -23
- package/src/gpu/kernels/conv2d.d.ts +0 -4
- package/src/gpu/kernels/conv2d_f16.wgsl +1 -1
- package/src/gpu/kernels/dequant.d.ts +0 -16
- package/src/gpu/kernels/dequant.js +1 -71
- package/src/gpu/kernels/dequant_mxfp4_expert_f16.wgsl +1 -1
- package/src/gpu/kernels/dequant_shared.wgsl +1 -55
- package/src/gpu/kernels/dispatch.d.ts +12 -97
- package/src/gpu/kernels/dispatch.js +0 -117
- package/src/gpu/kernels/energy_eval_f16.wgsl +1 -1
- package/src/gpu/kernels/energy_quintel_grad_f16.wgsl +1 -1
- package/src/gpu/kernels/energy_quintel_reduce_f16.wgsl +1 -1
- package/src/gpu/kernels/energy_quintel_update_f16.wgsl +1 -1
- package/src/gpu/kernels/energy_update_f16.wgsl +1 -1
- package/src/gpu/kernels/fused-gate-up-gelu.js +121 -0
- package/src/gpu/kernels/fused_ffn_f16.wgsl +1 -1
- package/src/gpu/kernels/fused_ffn_q4k.wgsl +0 -19
- package/src/gpu/kernels/fused_ffn_q4k_f16.wgsl +0 -18
- package/src/gpu/kernels/fused_gate_up_gelu_gated_f16.wgsl +124 -0
- package/src/gpu/kernels/fused_matmul_q4.wgsl +23 -439
- package/src/gpu/kernels/fused_matmul_q4_batched_f16.wgsl +1 -1
- package/src/gpu/kernels/fused_matmul_q4_batched_f16a.wgsl +0 -1
- package/src/gpu/kernels/fused_matmul_q4_batched_f16acc_f16a.wgsl +158 -0
- package/src/gpu/kernels/fused_matmul_q4_f16a.wgsl +0 -1
- package/src/gpu/kernels/fused_matmul_q4_multicol_f16.wgsl +0 -1
- package/src/gpu/kernels/fused_matmul_q4_multicol_f16a.wgsl +32 -33
- package/src/gpu/kernels/fused_matmul_q4_multicol_f16a_f32acc.wgsl +209 -0
- package/src/gpu/kernels/fused_matmul_q4_prefill_tiled_f16.wgsl +261 -0
- package/src/gpu/kernels/fused_matmul_q4_widetile.wgsl +164 -0
- package/src/gpu/kernels/fused_matmul_q4_widetile_f16.wgsl +198 -0
- package/src/gpu/kernels/fused_matmul_q4_widetile_f16a.wgsl +161 -0
- package/src/gpu/kernels/fused_matmul_q4_widetile_residual.wgsl +173 -0
- package/src/gpu/kernels/fused_matmul_rmsnorm_f16.wgsl +1 -1
- package/src/gpu/kernels/fused_rmsnorm_q4_widetile.wgsl +271 -0
- package/src/gpu/kernels/fused_swiglu_f16.wgsl +1 -1
- package/src/gpu/kernels/gather.d.ts +0 -7
- package/src/gpu/kernels/gather_f16.wgsl +1 -1
- package/src/gpu/kernels/gelu_f16.wgsl +14 -15
- package/src/gpu/kernels/groupnorm.d.ts +0 -4
- package/src/gpu/kernels/groupnorm_apply_f16.wgsl +1 -1
- package/src/gpu/kernels/groupnorm_stats_f16.wgsl +1 -1
- package/src/gpu/kernels/index.d.ts +6 -5
- package/src/gpu/kernels/index.js +3 -4
- package/src/gpu/kernels/kv_quantize_turboquant.wgsl +0 -2
- package/src/gpu/kernels/kv_quantize_turboquant_prod.wgsl +0 -2
- package/src/gpu/kernels/layernorm.d.ts +0 -4
- package/src/gpu/kernels/layernorm_f16.wgsl +1 -1
- package/src/gpu/kernels/{sana_linear_attention.d.ts → linear_attention.d.ts} +5 -5
- package/src/gpu/kernels/{sana_linear_attention.js → linear_attention.js} +16 -16
- package/src/gpu/kernels/logit-merge.d.ts +0 -24
- package/src/gpu/kernels/logit-merge.js +1 -17
- package/src/gpu/kernels/matmul-dispatch.js +7 -1
- package/src/gpu/kernels/matmul-selection.js +68 -1
- package/src/gpu/kernels/matmul.d.ts +0 -28
- package/src/gpu/kernels/matmul.js +101 -7
- package/src/gpu/kernels/matmul_f16_tiled.wgsl +33 -34
- package/src/gpu/kernels/matmul_gemv_residual_f16.wgsl +1 -1
- package/src/gpu/kernels/matmul_gemv_subgroup.wgsl +0 -1
- package/src/gpu/kernels/matmul_gemv_subgroup_f16a.wgsl +24 -25
- package/src/gpu/kernels/modulate_f16.wgsl +1 -1
- package/src/gpu/kernels/moe.d.ts +0 -56
- package/src/gpu/kernels/moe.js +0 -15
- package/src/gpu/kernels/moe_gather_f16.wgsl +1 -1
- package/src/gpu/kernels/pixel_shuffle_f16.wgsl +1 -1
- package/src/gpu/kernels/residual.d.ts +0 -12
- package/src/gpu/kernels/residual_f16.wgsl +2 -2
- package/src/gpu/kernels/residual_f16_vec4.wgsl +5 -5
- package/src/gpu/kernels/rmsnorm.d.ts +0 -8
- package/src/gpu/kernels/rmsnorm.js +17 -1
- package/src/gpu/kernels/rmsnorm.wgsl +5 -6
- package/src/gpu/kernels/rmsnorm_f16.wgsl +65 -39
- package/src/gpu/kernels/rope.d.ts +1 -5
- package/src/gpu/kernels/rope.js +12 -2
- package/src/gpu/kernels/rope.wgsl +33 -19
- package/src/gpu/kernels/rope_f16.wgsl +39 -25
- package/src/gpu/kernels/sample.d.ts +0 -5
- package/src/gpu/kernels/sample.wgsl +7 -6
- package/src/gpu/kernels/sample_f16.wgsl +28 -30
- package/src/gpu/kernels/scatter_add_dynamic_f16.wgsl +1 -1
- package/src/gpu/kernels/shader-cache.d.ts +13 -0
- package/src/gpu/kernels/shader-cache.js +27 -0
- package/src/gpu/kernels/silu_f16.wgsl +1 -1
- package/src/gpu/kernels/softmax.wgsl +4 -144
- package/src/gpu/kernels/split_qg_f16.wgsl +1 -1
- package/src/gpu/kernels/split_qkv_f16.wgsl +1 -1
- package/src/gpu/kernels/topk_f16.wgsl +1 -1
- package/src/gpu/kernels/turboquant-codebook.js +11 -40
- package/src/gpu/kernels/uniform-utils.d.ts +18 -0
- package/src/gpu/kernels/upsample2d.d.ts +0 -4
- package/src/gpu/kernels/upsample2d_f16.wgsl +1 -1
- package/src/gpu/kernels/utils.d.ts +28 -0
- package/src/gpu/kernels/utils.js +41 -0
- package/src/gpu/perf-guards.d.ts +0 -12
- package/src/gpu/perf-guards.js +0 -62
- package/src/gpu/profiler.d.ts +0 -21
- package/src/gpu/profiler.js +0 -34
- package/src/gpu/submit-tracker.d.ts +0 -57
- package/src/gpu/submit-tracker.js +3 -62
- package/src/gpu/tensor.d.ts +0 -23
- package/src/gpu/tensor.js +0 -41
- package/src/gpu/uniform-cache.d.ts +0 -11
- package/src/gpu/uniform-cache.js +0 -23
- package/src/index.d.ts +1 -0
- package/src/index.js +1 -0
- package/src/inference/browser-harness-model-helpers.d.ts +0 -9
- package/src/inference/browser-harness-model-helpers.js +22 -13
- package/src/inference/browser-harness-runtime-helpers.d.ts +0 -5
- package/src/inference/browser-harness-runtime-helpers.js +2 -2
- package/src/inference/browser-harness-suite-helpers.d.ts +4 -1
- package/src/inference/browser-harness-suite-helpers.js +87 -1
- package/src/inference/browser-harness-text-helpers.d.ts +12 -0
- package/src/inference/browser-harness-text-helpers.js +223 -14
- package/src/inference/browser-harness.d.ts +0 -29
- package/src/inference/browser-harness.js +264 -71
- package/src/inference/kv-cache/base.d.ts +6 -0
- package/src/inference/kv-cache/base.js +16 -1
- package/src/inference/kv-cache/mixed-geometry.d.ts +1 -0
- package/src/inference/kv-cache/mixed-geometry.js +12 -2
- package/src/inference/kv-cache/quantized.js +2 -0
- package/src/inference/kv-cache/tiered.js +7 -0
- package/src/inference/kv-cache/types.d.ts +2 -0
- package/src/inference/multi-model-network.d.ts +1 -1
- package/src/inference/multi-model-network.js +1 -2
- package/src/inference/network-evolution.d.ts +0 -12
- package/src/inference/network-evolution.js +2 -2
- package/src/inference/pipelines/context.js +17 -0
- package/src/inference/pipelines/diffusion/init.js +2 -2
- package/src/inference/pipelines/diffusion/pipeline.js +41 -96
- package/src/inference/pipelines/diffusion/text-encoder-gpu.d.ts +0 -5
- package/src/inference/pipelines/diffusion/text-encoder-gpu.js +0 -283
- package/src/inference/pipelines/diffusion/vae.js +4 -4
- package/src/inference/pipelines/energy/quintel.d.ts +15 -0
- package/src/inference/pipelines/structured/index.d.ts +2 -0
- package/src/inference/pipelines/structured/index.js +2 -0
- package/src/inference/pipelines/structured/json-grammar-mask.js +143 -0
- package/src/inference/pipelines/structured/tsv-grammar-mask.js +166 -0
- package/src/inference/pipelines/text/attention/attn-config.d.ts +0 -5
- package/src/inference/pipelines/text/attention/attn-config.js +1 -1
- package/src/inference/pipelines/text/attention/projections.d.ts +2 -0
- package/src/inference/pipelines/text/attention/projections.js +70 -23
- package/src/inference/pipelines/text/attention/record.js +118 -5
- package/src/inference/pipelines/text/attention/rmsnorm-fusion-gate.d.ts +8 -0
- package/src/inference/pipelines/text/attention/rmsnorm-fusion-gate.js +19 -0
- package/src/inference/pipelines/text/attention/run.js +97 -6
- package/src/inference/pipelines/text/attention/types.d.ts +2 -0
- package/src/inference/pipelines/text/chat-format.d.ts +0 -10
- package/src/inference/pipelines/text/chat-format.js +1 -8
- package/src/inference/pipelines/text/config.d.ts +3 -25
- package/src/inference/pipelines/text/config.js +105 -3
- package/src/inference/pipelines/text/debug-utils/utils.js +2 -23
- package/src/inference/pipelines/text/drift-policy.js +1 -1
- package/src/inference/pipelines/text/embed.d.ts +22 -27
- package/src/inference/pipelines/text/embed.js +0 -39
- package/src/inference/pipelines/text/execution-plan.d.ts +5 -3
- package/src/inference/pipelines/text/execution-plan.js +15 -2
- package/src/inference/pipelines/text/execution-runtime-builders.d.ts +0 -4
- package/src/inference/pipelines/text/execution-runtime-builders.js +0 -11
- package/src/inference/pipelines/text/execution-v1.d.ts +18 -0
- package/src/inference/pipelines/text/execution-v1.js +94 -25
- package/src/inference/pipelines/text/ffn/dense.js +180 -3
- package/src/inference/pipelines/text/ffn/standard.js +40 -13
- package/src/inference/pipelines/text/generator-decode-policy.js +243 -0
- package/src/inference/pipelines/text/generator-helpers.d.ts +0 -2
- package/src/inference/pipelines/text/generator-helpers.js +4 -5
- package/src/inference/pipelines/text/generator-prefill-helpers.js +293 -0
- package/src/inference/pipelines/text/generator-runtime.d.ts +0 -7
- package/src/inference/pipelines/text/generator-runtime.js +8 -6
- package/src/inference/pipelines/text/generator-steps.d.ts +22 -7
- package/src/inference/pipelines/text/generator-steps.js +8 -3
- package/src/inference/pipelines/text/generator.d.ts +0 -6
- package/src/inference/pipelines/text/generator.js +223 -547
- package/src/inference/pipelines/text/init-chat-templates.js +86 -0
- package/src/inference/pipelines/text/init-qkv-fusion.js +201 -0
- package/src/inference/pipelines/text/init-rope.js +408 -0
- package/src/inference/pipelines/text/init.d.ts +5 -2
- package/src/inference/pipelines/text/init.js +183 -724
- package/src/inference/pipelines/text/kernel-trace.d.ts +0 -33
- package/src/inference/pipelines/text/kernel-trace.js +0 -33
- package/src/inference/pipelines/text/layer-plan-gpu.js +542 -0
- package/src/inference/pipelines/text/layer-plan.d.ts +0 -5
- package/src/inference/pipelines/text/layer-plan.js +1 -1
- package/src/inference/pipelines/text/layer.d.ts +70 -8
- package/src/inference/pipelines/text/layer.js +105 -537
- package/src/inference/pipelines/text/linear-attention.d.ts +2 -1
- package/src/inference/pipelines/text/load-errors.d.ts +21 -0
- package/src/inference/pipelines/text/load-errors.js +59 -0
- package/src/inference/pipelines/text/logits/cpu.js +2 -20
- package/src/inference/pipelines/text/logits/gpu.js +2 -0
- package/src/inference/pipelines/text/logits/index.d.ts +1 -1
- package/src/inference/pipelines/text/model-load.d.ts +5 -4
- package/src/inference/pipelines/text/model-load.js +74 -17
- package/src/inference/pipelines/text/ops.d.ts +13 -2
- package/src/inference/pipelines/text/ops.js +1 -1
- package/src/inference/pipelines/text/per-layer-inputs.js +31 -6
- package/src/inference/pipelines/text/probes.d.ts +3 -6
- package/src/inference/pipelines/text/probes.js +120 -31
- package/src/inference/pipelines/text/sampling.d.ts +0 -5
- package/src/inference/pipelines/text/sampling.js +140 -0
- package/src/inference/pipelines/text/stage-names.js +3 -0
- package/src/inference/pipelines/text/state.js +2 -0
- package/src/inference/pipelines/text/tsir-fixture-writer.js +409 -0
- package/src/inference/pipelines/text/types.d.ts +8 -78
- package/src/inference/pipelines/text/weights.d.ts +0 -50
- package/src/inference/pipelines/text/weights.js +2 -51
- package/src/inference/pipelines/text.d.ts +9 -1
- package/src/inference/pipelines/text.js +223 -27
- package/src/inference/pipelines/vision/index.js +0 -107
- package/src/inference/runtime-model.d.ts +0 -7
- package/src/inference/runtime-model.js +0 -23
- package/src/inference/test-harness.d.ts +6 -22
- package/src/inference/test-harness.js +97 -23
- package/src/inference/tokenizer.js +1 -10
- package/src/inference/tokenizers/sentencepiece.js +65 -15
- package/src/inference/tokenizers/types.d.ts +0 -21
- package/src/loader/doppler-loader.d.ts +3 -0
- package/src/loader/doppler-loader.js +90 -12
- package/src/loader/embedding-loader.js +13 -3
- package/src/loader/loader-types.d.ts +2 -0
- package/src/loader/manifest-config.js +25 -1
- package/src/loader/memory-monitor.d.ts +17 -5
- package/src/loader/quantization-constants.d.ts +6 -0
- package/src/loader/shard-resolver.d.ts +1 -0
- package/src/loader/shard-resolver.js +15 -0
- package/src/loader/tensors/source-transform.js +6 -1
- package/src/loader/tensors/tensor-loader.d.ts +2 -9
- package/src/loader/tensors/tensor-loader.js +84 -1
- package/src/memory/address-table.d.ts +1 -9
- package/src/memory/buffer-pool.d.ts +0 -16
- package/src/memory/buffer-pool.js +1 -33
- package/src/memory/heap-manager.d.ts +18 -43
- package/src/memory/unified-detect.d.ts +19 -34
- package/src/models/embeddinggemma.d.ts +7 -0
- package/src/models/embeddinggemma.js +24 -0
- package/src/models/family.d.ts +24 -0
- package/src/models/family.js +25 -0
- package/src/models/gemma3.d.ts +7 -0
- package/src/models/gemma3.js +33 -0
- package/src/models/gemma4.d.ts +7 -0
- package/src/models/gemma4.js +33 -0
- package/src/models/qwen3.d.ts +7 -0
- package/src/models/qwen3.js +36 -0
- package/src/rules/converter/tokenizer.rules.json +5 -0
- package/src/rules/execution-rules-contract-check.js +72 -3
- package/src/rules/inference/capability-transforms.rules.json +87 -0
- package/src/rules/inference/execution.rules.json +11 -1
- package/src/rules/inference/ffn.rules.json +6 -2
- package/src/rules/loader/tensor-loader.rules.json +9 -0
- package/src/rules/rule-registry.js +53 -75
- package/src/storage/artifact-storage-context.d.ts +11 -3
- package/src/storage/artifact-storage-context.js +231 -26
- package/src/storage/backends/opfs-store.js +138 -88
- package/src/storage/blake3.js +9 -1
- package/src/storage/export.js +1 -5
- package/src/storage/normalize-model-id.js +3 -0
- package/src/storage/preflight.js +11 -0
- package/src/storage/quickstart-downloader.js +44 -67
- package/src/storage/reports.js +35 -11
- package/src/storage/shard-manager.d.ts +8 -2
- package/src/storage/shard-manager.js +176 -15
- package/src/storage/source-artifact-store.d.ts +0 -2
- package/src/storage/source-artifact-store.js +3 -9
- package/src/sw.js +0 -3
- package/src/tooling/browser-command-runner.js +39 -0
- package/src/tooling/command-api-constants.js +1 -1
- package/src/tooling/command-api-family-normalizers.d.ts +1 -0
- package/src/tooling/command-api-family-normalizers.js +49 -7
- package/src/tooling/command-api-helpers.d.ts +0 -2
- package/src/tooling/command-api-helpers.js +6 -2
- package/src/tooling/command-api.d.ts +20 -1
- package/src/tooling/command-api.js +10 -1
- package/src/tooling/command-envelope.d.ts +0 -1
- package/src/tooling/command-envelope.js +1 -1
- package/src/tooling/command-runner-shared.d.ts +10 -0
- package/src/tooling/conversion-config-materializer.d.ts +0 -10
- package/src/tooling/conversion-config-materializer.js +1 -1
- package/src/tooling/diagnose-runner.js +1 -4
- package/src/tooling/hf-registry-utils.d.ts +2 -6
- package/src/tooling/hf-registry-utils.js +90 -19
- package/src/tooling/kernel-path-builder/index.js +1 -4
- package/src/tooling/kernel-path-builder/runtime-overlay.js +1 -3
- package/src/tooling/lean-execution-contract-runner.d.ts +0 -10
- package/src/tooling/lean-execution-contract-runner.js +2 -2
- package/src/tooling/litert-package-runtime.d.ts +1 -1
- package/src/tooling/litert-package-runtime.js +21 -24
- package/src/tooling/node-browser-command-runner.d.ts +15 -0
- package/src/tooling/node-browser-command-runner.js +97 -13
- package/src/tooling/node-command-runner.js +36 -0
- package/src/tooling/node-convert-worker-pool.js +12 -0
- package/src/tooling/node-convert-worker.js +10 -1
- package/src/tooling/node-converter.js +90 -1
- package/src/tooling/node-source-runtime.js +1 -10
- package/src/tooling/opfs-cache.js +1 -4
- package/src/tooling/precision-replay-math.js +174 -0
- package/src/tooling/program-bundle-host.d.ts +16 -0
- package/src/tooling/program-bundle-host.js +20 -0
- package/src/tooling/program-bundle-parity.d.ts +28 -0
- package/src/tooling/program-bundle-parity.js +210 -0
- package/src/tooling/program-bundle.d.ts +61 -0
- package/src/tooling/program-bundle.js +1119 -0
- package/src/tooling/rdrr-integrity-refresh.d.ts +32 -0
- package/src/tooling/rdrr-integrity-refresh.js +271 -0
- package/src/tooling/reference-verify.js +191 -0
- package/src/tooling/source-artifact-adapter.d.ts +0 -20
- package/src/tooling/source-artifact-adapter.js +4 -4
- package/src/tooling/source-package-profiles.js +1 -7
- package/src/tooling/source-runtime-bundle.d.ts +2 -0
- package/src/tooling/source-runtime-bundle.js +3 -22
- package/src/tooling/source-runtime-converter-config.d.ts +0 -5
- package/src/tooling/source-runtime-converter-config.js +13 -9
- package/src/tooling/source-runtime-materializer.js +1 -7
- package/src/tooling-exports/device.d.ts +11 -0
- package/src/tooling-exports/device.js +14 -0
- package/src/tooling-exports/manifest.d.ts +8 -0
- package/src/tooling-exports/manifest.js +11 -0
- package/src/tooling-exports/storage.d.ts +25 -0
- package/src/tooling-exports/storage.js +29 -0
- package/src/tooling-exports/structured.d.ts +1 -0
- package/src/tooling-exports/structured.js +5 -0
- package/src/tooling-exports.d.ts +23 -0
- package/src/tooling-exports.js +17 -0
- package/src/tooling-exports.shared.d.ts +27 -30
- package/src/tooling-exports.shared.js +28 -29
- package/src/types/index.d.ts +0 -3
- package/src/types/index.js +0 -3
- package/src/utils/array-buffer.d.ts +4 -0
- package/src/utils/array-buffer.js +9 -0
- package/src/utils/canonical-hash.d.ts +7 -0
- package/src/utils/canonical-hash.js +74 -0
- package/src/utils/clone-json.d.ts +1 -0
- package/src/utils/clone-json.js +7 -0
- package/src/utils/encode-utf8.js +5 -0
- package/src/utils/index.d.ts +4 -1
- package/src/utils/index.js +4 -1
- package/src/utils/runtime-env.js +7 -0
- package/src/utils/sha256.d.ts +1 -0
- package/src/utils/sha256.js +18 -4
- package/src/utils/stable-sort-object.d.ts +1 -0
- package/src/utils/stable-sort-object.js +13 -0
- package/src/version.js +1 -1
- package/tools/convert-safetensors-node.js +1 -2
- package/src/bootstrap.d.ts +0 -1
- package/src/config/conversion/sana/sana-sprint-0.6b-f16.json +0 -247
- package/src/config/kernels/kernel-ref.d.ts +0 -17
- package/src/config/kernels/kernel-ref.js +0 -75
- package/src/config/runtime/kernels/dequant-f16-q4k.json +0 -13
- package/src/config/runtime/kernels/dequant-f32-q4k.json +0 -13
- package/src/config/runtime/kernels/embeddinggemma-q4k-dequant-f32a.json +0 -37
- package/src/config/runtime/kernels/fused-q4k.json +0 -13
- package/src/config/runtime/kernels/gemma2-q4k-dequant-f16a.json +0 -33
- package/src/config/runtime/kernels/gemma2-q4k-dequant-f32a-nosubgroups.json +0 -33
- package/src/config/runtime/kernels/gemma2-q4k-fused-f32a.json +0 -33
- package/src/config/runtime/kernels/safe-q4k.json +0 -13
- package/src/config/runtime/overlays/debug-f32a.json +0 -3
- package/src/config/runtime/overlays/trace-config.json +0 -28
- package/src/config/runtime/profiles/gemma4-e2b-throughput.json +0 -39
- package/src/debug/perf.d.ts +0 -33
- package/src/debug/perf.js +0 -51
- package/src/errors/index.d.ts +0 -1
- package/src/errors/index.js +0 -1
- package/src/formats/gguf/index.d.ts +0 -8
- package/src/formats/gguf/index.js +0 -4
- package/src/formats/litert/index.d.ts +0 -1
- package/src/formats/litert/index.js +0 -1
- package/src/formats/safetensors/index.d.ts +0 -8
- package/src/formats/safetensors/index.js +0 -4
- package/src/formats/tflite/index.d.ts +0 -1
- package/src/formats/tflite/index.js +0 -1
- package/src/gpu/kernels/attention_decode_optimized.wgsl +0 -445
- package/src/gpu/kernels/dequant_mxfp4_vec4.wgsl +0 -116
- package/src/gpu/kernels/dequant_q8_0.wgsl +0 -98
- package/src/gpu/kernels/dequant_shared_vec4.wgsl +0 -160
- package/src/gpu/kernels/moe_gather_vec4.wgsl +0 -74
- package/src/inference/functiongemma.d.ts +0 -15
- package/src/inference/functiongemma.js +0 -1
- package/src/inference/pipelines/diffusion/sana-transformer.d.ts +0 -53
- package/src/inference/pipelines/diffusion/sana-transformer.js +0 -738
- package/src/rules/kernels/kernel-validator.d.ts +0 -24
- package/src/rules/kernels/kernel-validator.js +0 -160
- package/src/sw.d.ts +0 -1
- package/src/tooling/bench-runner.js +0 -64
- package/src/tooling/distill-runner.js +0 -56
- package/src/tooling/lora-runner.js +0 -37
- package/src/tooling/verify-runner.js +0 -63
- package/src/types/gpu.d.ts +0 -185
- package/src/types/gpu.js +0 -5
- package/src/types/inference.d.ts +0 -197
- package/src/types/inference.js +0 -5
- package/src/types/model.d.ts +0 -130
- package/src/types/model.js +0 -5
- package/src/utils/format-bytes.js +0 -20
- /package/src/config/{runtime/diagnostics → diagnostics}/drift-policies.json +0 -0
- /package/src/gpu/kernels/{sana_linear_attention_apply.wgsl → linear_attention_apply.wgsl} +0 -0
- /package/src/gpu/kernels/{sana_linear_attention_apply_f16.wgsl → linear_attention_apply_f16.wgsl} +0 -0
- /package/src/gpu/kernels/{sana_linear_attention_summary.wgsl → linear_attention_summary.wgsl} +0 -0
- /package/src/gpu/kernels/{sana_linear_attention_summary_f16.wgsl → linear_attention_summary_f16.wgsl} +0 -0
- /package/src/rules/kernels/{sana-linear-attention.rules.json → linear-attention.rules.json} +0 -0
|
@@ -1,35 +1,33 @@
|
|
|
1
|
-
---
|
|
2
|
-
+++
|
|
3
|
-
@@ -1,
|
|
1
|
+
--- src/gpu/kernels/rmsnorm.wgsl
|
|
2
|
+
+++ src/gpu/kernels/rmsnorm_f16.wgsl
|
|
3
|
+
@@ -1,46 +1,24 @@
|
|
4
4
|
-// RMSNorm Kernel with Fused Residual Add
|
|
5
|
-
+// AUTO-GENERATED from src/gpu/kernels/rmsnorm.wgsl.
|
|
6
|
-
+// Edit the source kernel and src/gpu/kernels/codegen/wgsl-variants.js, then run `npm run kernels:generate`.
|
|
7
|
-
+// RMSNorm Kernel with F16 Input/Output
|
|
8
5
|
//
|
|
9
6
|
-// RMSNorm(x) = x / sqrt(mean(x^2) + eps) * weight
|
|
10
|
-
|
|
11
|
-
+// Intermediate computations (sum of squares, RMS) remain in F32 for precision.
|
|
12
|
-
+// Weight buffer may be F16 or F32 (small size, precision matters).
|
|
13
|
-
//
|
|
7
|
+
-//
|
|
14
8
|
-// Optionally fuses residual addition (POST-norm, for Gemma 3 sandwich norm):
|
|
15
9
|
-// output = residual + RMSNorm(x) * weight
|
|
16
10
|
-//
|
|
17
11
|
-// Uses workgroup reduction for efficient mean calculation.
|
|
18
12
|
-// Subgroup variants use subgroupAdd() for 3-5x faster reductions.
|
|
19
|
-
|
|
13
|
+
+// F16 variant for reduced memory bandwidth when using F16 activations.
|
|
14
|
+
+// The all-f16 lane keeps normalization arithmetic in f16.
|
|
15
|
+
+// Weight buffer may be F16 or F32.
|
|
16
|
+
//
|
|
20
17
|
-// Uses override constants for compile-time feature selection:
|
|
21
18
|
-// - RMS_NORM_OFFSET: Use (1 + weight) pattern (Gemma models)
|
|
22
|
-
-// - HAS_RESIDUAL:
|
|
19
|
+
-// - HAS_RESIDUAL: Compile-time override (reserved for codegen patches, not used by JS dispatch)
|
|
23
20
|
-// - OUTPUT_PRENORM: Output pre-normalized values (before weight multiplication)
|
|
24
|
-
|
|
25
|
-
|
|
21
|
+
-
|
|
26
22
|
-enable subgroups;
|
|
27
|
-
|
|
28
|
-
|
|
23
|
+
-
|
|
29
24
|
-// =============================================================================
|
|
30
25
|
-// Override Constants (compile-time configuration)
|
|
31
26
|
-// =============================================================================
|
|
32
|
-
|
|
27
|
+
+// RMSNorm(x) = x / sqrt(mean(x^2) + eps) * weight
|
|
28
|
+
+
|
|
29
|
+
+enable f16;
|
|
30
|
+
|
|
33
31
|
override WORKGROUP_SIZE: u32 = 256u;
|
|
34
32
|
const MAX_WORKGROUP_SIZE: u32 = 256u;
|
|
35
33
|
-
|
|
@@ -39,8 +37,9 @@
|
|
|
39
37
|
override PRE_RESIDUAL: bool = false; // Add residual BEFORE normalization: rmsnorm(input + residual)
|
|
40
38
|
override OUTPUT_PRENORM: bool = false; // Write pre-norm sum to residual_sum_output binding
|
|
41
39
|
override WEIGHT_IS_F16: bool = false; // Weight buffer packed as f16 pairs
|
|
42
|
-
|
|
40
|
+
|
|
43
41
|
-const MAX_SUBGROUPS: u32 = 32u; // Support up to 32 subgroups per workgroup
|
|
42
|
+
-const MAX_CACHE_SIZE: u32 = 4608u; // Supports hiddenSize up to 1152 with 4 elements/thread
|
|
44
43
|
-
|
|
45
44
|
-// =============================================================================
|
|
46
45
|
-// Uniforms (per-dispatch)
|
|
@@ -56,77 +55,81 @@
|
|
|
56
55
|
token_stride: u32, // Workgroup rows per dispatch row
|
|
57
56
|
_pad0: u32,
|
|
58
57
|
_pad1: u32,
|
|
59
|
-
@@ -47
|
|
58
|
+
@@ -48,47 +26,32 @@
|
|
60
59
|
}
|
|
61
|
-
|
|
60
|
+
|
|
62
61
|
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
63
62
|
-@group(0) @binding(1) var<storage, read> input: array<f32>;
|
|
64
63
|
-@group(0) @binding(2) var<storage, read> weight: array<u32>; // [size] as f32 or packed f16
|
|
65
64
|
-@group(0) @binding(3) var<storage, read_write> output: array<f32>;
|
|
66
65
|
-@group(0) @binding(4) var<storage, read> residual: array<f32>; // Optional residual input
|
|
67
66
|
-@group(0) @binding(5) var<storage, read_write> residual_sum_output: array<f32>; // Pre-norm sum output (PRE_RESIDUAL + OUTPUT_PRENORM)
|
|
68
|
-
|
|
69
|
-
+@group(0) @binding(2) var<storage, read> weight: array<u32>; // F32 or packed F16
|
|
70
|
-
+@group(0) @binding(3) var<storage, read_write> output: array<f16>;
|
|
71
|
-
+@group(0) @binding(4) var<storage, read> residual: array<f16>; // Optional residual
|
|
72
|
-
+@group(0) @binding(5) var<storage, read_write> residual_sum_output: array<f16>; // Pre-norm sum output (PRE_RESIDUAL + OUTPUT_PRENORM)
|
|
73
|
-
|
|
67
|
+
-
|
|
74
68
|
-// =============================================================================
|
|
75
69
|
-// Shared Memory
|
|
76
70
|
-// =============================================================================
|
|
77
71
|
-
|
|
78
|
-
|
|
79
|
-
|
|
80
|
-
|
|
72
|
+
-var<workgroup> shared_sum: array<f32, MAX_WORKGROUP_SIZE>;
|
|
73
|
+
-
|
|
81
74
|
-// =============================================================================
|
|
82
75
|
-// Helper Functions
|
|
83
76
|
-// =============================================================================
|
|
84
|
-
|
|
77
|
+
+@group(0) @binding(1) var<storage, read> input: array<f16>;
|
|
78
|
+
+@group(0) @binding(2) var<storage, read> weight: array<u32>; // F32 or packed F16
|
|
79
|
+
+@group(0) @binding(3) var<storage, read_write> output: array<f16>;
|
|
80
|
+
+@group(0) @binding(4) var<storage, read> residual: array<f16>; // Optional residual
|
|
81
|
+
+@group(0) @binding(5) var<storage, read_write> residual_sum_output: array<f16>; // Pre-norm sum output (PRE_RESIDUAL + OUTPUT_PRENORM)
|
|
82
|
+
|
|
85
83
|
-// Apply weight with optional offset (Gemma uses 1 + weight pattern)
|
|
86
|
-
|
|
84
|
+
-fn apply_weight(w: f32) -> f32 {
|
|
85
|
+
+var<workgroup> shared_sum: array<f16, MAX_WORKGROUP_SIZE>;
|
|
86
|
+
+
|
|
87
|
+
+fn apply_weight(w: f16) -> f16 {
|
|
87
88
|
if (RMS_NORM_OFFSET) {
|
|
88
|
-
|
|
89
|
+
- return 1.0 + w;
|
|
89
90
|
- } else {
|
|
90
91
|
- return w;
|
|
92
|
+
+ return f16(1.0) + w;
|
|
91
93
|
}
|
|
92
94
|
+ return w;
|
|
93
95
|
}
|
|
94
|
-
|
|
95
|
-
|
|
96
|
-
|
|
97
|
-
|
|
98
|
-
|
|
99
|
-
|
|
100
|
-
|
|
101
|
-
|
|
102
|
-
|
|
103
|
-
-
|
|
104
|
-
- return HAS_RESIDUAL || (u.has_residual != 0u);
|
|
96
|
+
|
|
97
|
+
-fn load_weight(idx: u32) -> f32 {
|
|
98
|
+
+fn load_weight(idx: u32) -> f16 {
|
|
99
|
+
if (WEIGHT_IS_F16) {
|
|
100
|
+
let packed = weight[idx >> 1u];
|
|
101
|
+
let pair = unpack2x16float(packed);
|
|
102
|
+
- return select(pair.x, pair.y, (idx & 1u) == 1u);
|
|
103
|
+
+ return f16(select(pair.x, pair.y, (idx & 1u) == 1u));
|
|
104
|
+
}
|
|
105
|
+
- return bitcast<f32>(weight[idx]);
|
|
105
106
|
-}
|
|
106
107
|
-
|
|
108
|
+
-// Check if residual should be added (runtime uniform only — JS dispatch controls this)
|
|
109
|
+
-fn should_add_residual() -> bool {
|
|
110
|
+
- return u.has_residual != 0u;
|
|
111
|
+
+ return f16(bitcast<f32>(weight[idx]));
|
|
112
|
+
}
|
|
113
|
+
|
|
107
114
|
// Load input value, fusing residual add when PRE_RESIDUAL is active
|
|
108
|
-
|
|
109
|
-
|
|
110
|
-
|
|
115
|
+
-fn load_input(base_offset: u32, idx: u32) -> f32 {
|
|
116
|
+
+fn load_input(base_offset: u32, idx: u32) -> f16 {
|
|
117
|
+
let x = input[base_offset + idx];
|
|
111
118
|
if (PRE_RESIDUAL) {
|
|
112
|
-
|
|
113
|
-
|
|
114
|
-
}
|
|
115
|
-
return x;
|
|
119
|
+
return x + residual[base_offset + idx];
|
|
120
|
+
@@ -97,7 +60,7 @@
|
|
116
121
|
}
|
|
117
|
-
|
|
122
|
+
|
|
118
123
|
// Write pre-norm sum when OUTPUT_PRENORM is active (for downstream residual reuse)
|
|
119
|
-
|
|
124
|
+
-fn write_prenorm(base_offset: u32, idx: u32, val: f32) {
|
|
125
|
+
+fn write_prenorm(base_offset: u32, idx: u32, val: f16) {
|
|
120
126
|
if (OUTPUT_PRENORM && PRE_RESIDUAL) {
|
|
121
|
-
|
|
122
|
-
+ residual_sum_output[base_offset + idx] = f16(val);
|
|
127
|
+
residual_sum_output[base_offset + idx] = val;
|
|
123
128
|
}
|
|
124
|
-
|
|
125
|
-
|
|
126
|
-
@@ -108,10 +74,6 @@
|
|
129
|
+
@@ -107,10 +70,6 @@
|
|
127
130
|
return wg_id.y * max(u.token_stride, 1u) + wg_id.x;
|
|
128
131
|
}
|
|
129
|
-
|
|
132
|
+
|
|
130
133
|
-// =============================================================================
|
|
131
134
|
-// Main Entry Point
|
|
132
135
|
-// =============================================================================
|
|
@@ -134,74 +137,111 @@
|
|
|
134
137
|
// Main RMSNorm kernel - one workgroup per token
|
|
135
138
|
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
136
139
|
fn main(
|
|
137
|
-
@@ -
|
|
138
|
-
|
|
140
|
+
@@ -128,151 +87,44 @@
|
|
141
|
+
|
|
139
142
|
let base_offset = token_idx * size;
|
|
140
|
-
|
|
143
|
+
|
|
141
144
|
- // Each thread computes partial sum of squares
|
|
142
|
-
|
|
143
|
-
|
|
144
|
-
var local_sum_sq: f32 = 0.0;
|
|
145
|
+
- // When PRE_RESIDUAL is active, sum is computed on (input + residual)
|
|
146
|
+
- var local_sum_sq: f32 = 0.0;
|
|
145
147
|
let elements_per_thread = (size + WORKGROUP_SIZE - 1u) / WORKGROUP_SIZE;
|
|
146
|
-
|
|
148
|
+
|
|
149
|
+
+ var local_max: f16 = f16(0.0);
|
|
150
|
+
for (var i: u32 = 0u; i < elements_per_thread; i = i + 1u) {
|
|
151
|
+
let idx = thread_idx * elements_per_thread + i;
|
|
152
|
+
if (idx < size) {
|
|
153
|
+
- let x = load_input(base_offset, idx);
|
|
154
|
+
- local_sum_sq = local_sum_sq + x * x;
|
|
155
|
+
+ local_max = max(local_max, abs(load_input(base_offset, idx)));
|
|
156
|
+
}
|
|
157
|
+
}
|
|
158
|
+
|
|
159
|
+
- // Store local sum for reduction
|
|
160
|
+
- shared_sum[thread_idx] = local_sum_sq;
|
|
161
|
+
+ shared_sum[thread_idx] = local_max;
|
|
162
|
+
workgroupBarrier();
|
|
163
|
+
|
|
164
|
+
- // Parallel reduction to compute total sum of squares
|
|
165
|
+
for (var stride: u32 = WORKGROUP_SIZE / 2u; stride > 0u; stride = stride >> 1u) {
|
|
166
|
+
if (thread_idx < stride) {
|
|
167
|
+
- shared_sum[thread_idx] = shared_sum[thread_idx] + shared_sum[thread_idx + stride];
|
|
168
|
+
+ shared_sum[thread_idx] = max(shared_sum[thread_idx], shared_sum[thread_idx + stride]);
|
|
169
|
+
}
|
|
147
170
|
workgroupBarrier();
|
|
148
171
|
}
|
|
149
|
-
|
|
172
|
+
|
|
150
173
|
- // Compute RMS
|
|
151
|
-
|
|
152
|
-
|
|
153
|
-
|
|
154
|
-
|
|
155
|
-
|
|
156
|
-
|
|
157
|
-
|
|
158
|
-
|
|
174
|
+
- let mean_sq = shared_sum[0] / f32(size);
|
|
175
|
+
- let rms = sqrt(mean_sq + u.eps);
|
|
176
|
+
- let inv_rms = 1.0 / rms;
|
|
177
|
+
-
|
|
178
|
+
+ let row_scale = select(f16(1.0), shared_sum[0], shared_sum[0] > f16(0.0));
|
|
179
|
+
workgroupBarrier();
|
|
180
|
+
|
|
181
|
+
- // Apply normalization and weight
|
|
182
|
+
- for (var i: u32 = 0u; i < elements_per_thread; i = i + 1u) {
|
|
183
|
+
- let idx = thread_idx * elements_per_thread + i;
|
|
184
|
+
- if (idx < size) {
|
|
185
|
+
- let x = load_input(base_offset, idx);
|
|
186
|
+
-
|
|
159
187
|
- // Normalize and scale (with optional weight offset for Gemma)
|
|
160
|
-
|
|
161
|
-
|
|
162
|
-
|
|
188
|
+
- var result = x * inv_rms * apply_weight(load_weight(idx));
|
|
189
|
+
-
|
|
163
190
|
- // Add residual AFTER normalization (Gemma 3 sandwich norm pattern)
|
|
164
191
|
- if (should_add_residual()) {
|
|
165
192
|
- result = result + residual[base_offset + idx];
|
|
166
|
-
|
|
167
|
-
|
|
168
|
-
|
|
169
|
-
|
|
170
|
-
|
|
171
|
-
// Write pre-norm sum for downstream residual reuse
|
|
172
|
-
write_prenorm(base_offset, idx, x);
|
|
173
|
-
|
|
193
|
+
- }
|
|
194
|
+
-
|
|
195
|
+
- // Write pre-norm sum for downstream residual reuse
|
|
196
|
+
- write_prenorm(base_offset, idx, x);
|
|
197
|
+
-
|
|
174
198
|
- output[base_offset + idx] = result;
|
|
175
|
-
|
|
176
|
-
|
|
177
|
-
|
|
178
|
-
|
|
179
|
-
}
|
|
180
|
-
|
|
199
|
+
- }
|
|
200
|
+
- }
|
|
201
|
+
-}
|
|
202
|
+
-
|
|
181
203
|
-// =============================================================================
|
|
182
204
|
-// Small Hidden Size Entry Point
|
|
183
205
|
-// =============================================================================
|
|
184
206
|
-
|
|
185
|
-
|
|
186
|
-
|
|
207
|
+
-// Optimized version for hidden size <= WORKGROUP_SIZE (single pass)
|
|
208
|
+
-@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
187
209
|
-fn main_small(
|
|
188
|
-
|
|
189
|
-
|
|
190
|
-
|
|
191
|
-
)
|
|
192
|
-
|
|
193
|
-
|
|
194
|
-
|
|
195
|
-
|
|
210
|
+
- @builtin(local_invocation_id) local_id: vec3<u32>,
|
|
211
|
+
- @builtin(workgroup_id) wg_id: vec3<u32>
|
|
212
|
+
-) {
|
|
213
|
+
- let token_idx = token_index(wg_id);
|
|
214
|
+
- let thread_idx = local_id.x;
|
|
215
|
+
- let size = u.size;
|
|
216
|
+
-
|
|
217
|
+
- if (token_idx >= u.num_tokens) {
|
|
218
|
+
- return;
|
|
219
|
+
- }
|
|
220
|
+
-
|
|
221
|
+
- let base_offset = token_idx * size;
|
|
222
|
+
-
|
|
223
|
+
- // Each thread handles one element (for size <= 256)
|
|
224
|
+
- var x: f32 = 0.0;
|
|
225
|
+
- if (thread_idx < size) {
|
|
226
|
+
- x = load_input(base_offset, thread_idx);
|
|
227
|
+
- }
|
|
228
|
+
-
|
|
196
229
|
- // Sum of squares
|
|
197
|
-
|
|
198
|
-
|
|
199
|
-
|
|
200
|
-
|
|
201
|
-
|
|
202
|
-
|
|
203
|
-
|
|
204
|
-
|
|
230
|
+
- shared_sum[thread_idx] = x * x;
|
|
231
|
+
- workgroupBarrier();
|
|
232
|
+
-
|
|
233
|
+
- // Parallel reduction
|
|
234
|
+
- for (var stride: u32 = WORKGROUP_SIZE / 2u; stride > 0u; stride = stride >> 1u) {
|
|
235
|
+
- if (thread_idx < stride && thread_idx + stride < size) {
|
|
236
|
+
- shared_sum[thread_idx] = shared_sum[thread_idx] + shared_sum[thread_idx + stride];
|
|
237
|
+
- }
|
|
238
|
+
- workgroupBarrier();
|
|
239
|
+
- }
|
|
240
|
+
-
|
|
241
|
+
- // Compute inverse RMS
|
|
242
|
+
- let mean_sq = shared_sum[0] / f32(size);
|
|
243
|
+
- let inv_rms = 1.0 / sqrt(mean_sq + u.eps);
|
|
244
|
+
-
|
|
205
245
|
- // Apply normalization
|
|
206
246
|
- if (thread_idx < size) {
|
|
207
247
|
- var result = x * inv_rms * apply_weight(load_weight(thread_idx));
|
|
@@ -219,7 +259,7 @@
|
|
|
219
259
|
-
|
|
220
260
|
-// OPTIMIZED: Caches input to avoid double loads
|
|
221
261
|
-// Uses shared memory to store input values between passes
|
|
222
|
-
-var<workgroup> shared_cache: array<f32,
|
|
262
|
+
-var<workgroup> shared_cache: array<f32, MAX_CACHE_SIZE>;
|
|
223
263
|
-
|
|
224
264
|
-@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
225
265
|
-fn main_cached(
|
|
@@ -239,30 +279,39 @@
|
|
|
239
279
|
-
|
|
240
280
|
- // First pass: cache input and compute sum of squares
|
|
241
281
|
- var local_sum_sq: f32 = 0.0;
|
|
242
|
-
|
|
243
|
-
|
|
244
|
-
|
|
282
|
+
+ var local_sum_sq: f16 = f16(0.0);
|
|
283
|
+
+ let inv_size = f16(1.0) / f16(size);
|
|
284
|
+
for (var i: u32 = 0u; i < elements_per_thread; i = i + 1u) {
|
|
285
|
+
let idx = thread_idx * elements_per_thread + i;
|
|
286
|
+
if (idx < size) {
|
|
245
287
|
- let x = load_input(base_offset, idx);
|
|
246
288
|
- shared_cache[idx] = x; // Cache for second pass (avoids reloading residual)
|
|
247
289
|
- local_sum_sq = local_sum_sq + x * x;
|
|
248
|
-
|
|
249
|
-
|
|
250
|
-
|
|
251
|
-
|
|
252
|
-
|
|
253
|
-
|
|
254
|
-
|
|
255
|
-
|
|
256
|
-
|
|
257
|
-
|
|
258
|
-
|
|
259
|
-
|
|
260
|
-
|
|
290
|
+
+ let x = load_input(base_offset, idx) / row_scale;
|
|
291
|
+
+ local_sum_sq = local_sum_sq + (x * x) * inv_size;
|
|
292
|
+
}
|
|
293
|
+
}
|
|
294
|
+
|
|
295
|
+
+ // Store local sum for reduction
|
|
296
|
+
shared_sum[thread_idx] = local_sum_sq;
|
|
297
|
+
workgroupBarrier();
|
|
298
|
+
|
|
299
|
+
+ // Parallel reduction to compute total sum of squares
|
|
300
|
+
for (var stride: u32 = WORKGROUP_SIZE / 2u; stride > 0u; stride = stride >> 1u) {
|
|
301
|
+
if (thread_idx < stride) {
|
|
302
|
+
shared_sum[thread_idx] = shared_sum[thread_idx] + shared_sum[thread_idx + stride];
|
|
303
|
+
@@ -280,117 +132,38 @@
|
|
304
|
+
workgroupBarrier();
|
|
305
|
+
}
|
|
306
|
+
|
|
261
307
|
- let mean_sq = shared_sum[0] / f32(size);
|
|
262
308
|
- let inv_rms = 1.0 / sqrt(mean_sq + u.eps);
|
|
263
|
-
|
|
264
|
-
|
|
265
|
-
|
|
309
|
+
+ let mean_sq = shared_sum[0];
|
|
310
|
+
+ let eps_scaled = f16(u.eps) / (row_scale * row_scale);
|
|
311
|
+
+ let inv_rms = (f16(1.0) / row_scale) / sqrt(mean_sq + eps_scaled);
|
|
312
|
+
|
|
313
|
+
workgroupBarrier();
|
|
314
|
+
|
|
266
315
|
- // Second pass: normalize cached input
|
|
267
316
|
- for (var i: u32 = 0u; i < elements_per_thread; i = i + 1u) {
|
|
268
317
|
- let idx = thread_idx * elements_per_thread + i;
|
|
@@ -344,87 +393,108 @@
|
|
|
344
393
|
- let mean_sq = total_sum / f32(size);
|
|
345
394
|
- let inv_rms = 1.0 / sqrt(mean_sq + u.eps);
|
|
346
395
|
-
|
|
347
|
-
|
|
348
|
-
|
|
349
|
-
|
|
350
|
-
|
|
351
|
-
|
|
352
|
-
|
|
353
|
-
|
|
396
|
+
// Apply normalization and weight
|
|
397
|
+
for (var i: u32 = 0u; i < elements_per_thread; i = i + 1u) {
|
|
398
|
+
let idx = thread_idx * elements_per_thread + i;
|
|
399
|
+
if (idx < size) {
|
|
400
|
+
let x = load_input(base_offset, idx);
|
|
401
|
+
+
|
|
402
|
+
var result = x * inv_rms * apply_weight(load_weight(idx));
|
|
403
|
+
|
|
354
404
|
- if (should_add_residual()) {
|
|
355
|
-
|
|
356
|
-
|
|
357
|
-
|
|
358
|
-
|
|
359
|
-
|
|
360
|
-
-
|
|
361
|
-
|
|
362
|
-
|
|
363
|
-
|
|
405
|
+
+ // Add residual AFTER normalization
|
|
406
|
+
+ if (u.has_residual == 1u) {
|
|
407
|
+
result = result + residual[base_offset + idx];
|
|
408
|
+
}
|
|
409
|
+
|
|
410
|
+
+ // Write pre-norm sum for downstream residual reuse
|
|
411
|
+
write_prenorm(base_offset, idx, x);
|
|
412
|
+
+
|
|
413
|
+
output[base_offset + idx] = result;
|
|
414
|
+
}
|
|
415
|
+
}
|
|
416
|
+
}
|
|
417
|
+
|
|
364
418
|
-// Subgroup-accelerated RMSNorm for small hidden sizes (size <= 256)
|
|
365
419
|
-// Each thread handles one element, uses subgroup reduction
|
|
366
|
-
|
|
420
|
+
+// Optimized version for hidden size <= WORKGROUP_SIZE (single pass)
|
|
421
|
+
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
367
422
|
-fn main_small_subgroup(
|
|
368
|
-
|
|
423
|
+
+fn rmsnorm_small_f16(
|
|
424
|
+
@builtin(local_invocation_id) local_id: vec3<u32>,
|
|
369
425
|
- @builtin(workgroup_id) wg_id: vec3<u32>,
|
|
370
426
|
- @builtin(subgroup_invocation_id) sg_lane: u32,
|
|
371
427
|
- @builtin(subgroup_size) sg_size: u32,
|
|
372
|
-
|
|
373
|
-
|
|
374
|
-
|
|
375
|
-
|
|
376
|
-
-
|
|
377
|
-
|
|
378
|
-
|
|
379
|
-
|
|
380
|
-
-
|
|
381
|
-
- let base_offset = token_idx * size;
|
|
428
|
+
+ @builtin(workgroup_id) wg_id: vec3<u32>
|
|
429
|
+
) {
|
|
430
|
+
let token_idx = token_index(wg_id);
|
|
431
|
+
let thread_idx = local_id.x;
|
|
432
|
+
@@ -401,45 +174,46 @@
|
|
433
|
+
}
|
|
434
|
+
|
|
435
|
+
let base_offset = token_idx * size;
|
|
382
436
|
- let subgroup_id = thread_idx / sg_size;
|
|
383
437
|
- let num_subgroups = (WORKGROUP_SIZE + sg_size - 1u) / sg_size;
|
|
384
|
-
|
|
438
|
+
|
|
385
439
|
- // Each thread handles one element
|
|
386
440
|
- var x: f32 = 0.0;
|
|
387
441
|
- var x_sq: f32 = 0.0;
|
|
388
|
-
+ //
|
|
442
|
+
+ // Each thread handles one element (for size <= 256)
|
|
443
|
+
+ var x: f16 = f16(0.0);
|
|
389
444
|
if (thread_idx < size) {
|
|
390
|
-
|
|
445
|
+
x = load_input(base_offset, thread_idx);
|
|
391
446
|
- x_sq = x * x;
|
|
392
|
-
|
|
393
|
-
|
|
447
|
+
}
|
|
448
|
+
|
|
394
449
|
- // Phase 1: Subgroup reduction
|
|
395
450
|
- let sg_sum = subgroupAdd(x_sq);
|
|
396
|
-
|
|
451
|
+
+ shared_sum[thread_idx] = abs(x);
|
|
452
|
+
+ workgroupBarrier();
|
|
453
|
+
|
|
397
454
|
- // Phase 2: Store to shared memory
|
|
398
455
|
- if (sg_lane == 0u && subgroup_id < num_subgroups) {
|
|
399
456
|
- sg_partial_sums[subgroup_id] = sg_sum;
|
|
400
|
-
|
|
401
|
-
|
|
402
|
-
|
|
457
|
+
+ for (var stride: u32 = WORKGROUP_SIZE / 2u; stride > 0u; stride = stride >> 1u) {
|
|
458
|
+
+ if (thread_idx < stride) {
|
|
459
|
+
+ shared_sum[thread_idx] = max(shared_sum[thread_idx], shared_sum[thread_idx + stride]);
|
|
460
|
+
+ }
|
|
461
|
+
+ workgroupBarrier();
|
|
462
|
+
}
|
|
463
|
+
+
|
|
464
|
+
+ let row_scale = select(f16(1.0), shared_sum[0], shared_sum[0] > f16(0.0));
|
|
465
|
+
+ let x_scaled = x / row_scale;
|
|
466
|
+
+ let inv_size = f16(1.0) / f16(size);
|
|
467
|
+
+ shared_sum[thread_idx] = (x_scaled * x_scaled) * inv_size;
|
|
468
|
+
workgroupBarrier();
|
|
469
|
+
|
|
403
470
|
- // Phase 3: Final reduction - thread 0 sums, then broadcast via shared memory
|
|
404
471
|
- if (thread_idx == 0u) {
|
|
405
472
|
- var sum: f32 = 0.0;
|
|
406
473
|
- for (var s = 0u; s < num_subgroups; s++) {
|
|
407
474
|
- sum += sg_partial_sums[s];
|
|
408
|
-
|
|
475
|
+
+ // Parallel reduction
|
|
476
|
+
+ for (var stride: u32 = WORKGROUP_SIZE / 2u; stride > 0u; stride = stride >> 1u) {
|
|
477
|
+
+ if (thread_idx < stride && thread_idx + stride < size) {
|
|
478
|
+
+ shared_sum[thread_idx] = shared_sum[thread_idx] + shared_sum[thread_idx + stride];
|
|
479
|
+
}
|
|
409
480
|
- sg_partial_sums[0] = sum;
|
|
410
|
-
|
|
481
|
+
+ workgroupBarrier();
|
|
482
|
+
}
|
|
411
483
|
- workgroupBarrier();
|
|
412
484
|
- let total_sum = sg_partial_sums[0];
|
|
413
|
-
|
|
414
|
-
|
|
485
|
+
|
|
486
|
+
// Compute inverse RMS
|
|
415
487
|
- let mean_sq = total_sum / f32(size);
|
|
416
488
|
- let inv_rms = 1.0 / sqrt(mean_sq + u.eps);
|
|
417
|
-
|
|
418
|
-
|
|
419
|
-
|
|
489
|
+
+ let mean_sq = shared_sum[0];
|
|
490
|
+
+ let eps_scaled = f16(u.eps) / (row_scale * row_scale);
|
|
491
|
+
+ let inv_rms = (f16(1.0) / row_scale) / sqrt(mean_sq + eps_scaled);
|
|
492
|
+
|
|
493
|
+
// Apply normalization
|
|
494
|
+
if (thread_idx < size) {
|
|
420
495
|
var result = x * inv_rms * apply_weight(load_weight(thread_idx));
|
|
421
496
|
- if (should_add_residual()) {
|
|
422
|
-
- result = result + residual[base_offset + thread_idx];
|
|
423
497
|
+ if (u.has_residual == 1u) {
|
|
424
|
-
|
|
498
|
+
result = result + residual[base_offset + thread_idx];
|
|
425
499
|
}
|
|
426
500
|
write_prenorm(base_offset, thread_idx, x);
|
|
427
|
-
- output[base_offset + thread_idx] = result;
|
|
428
|
-
+ output[base_offset + thread_idx] = f16(result);
|
|
429
|
-
}
|
|
430
|
-
}
|