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
|
@@ -0,0 +1,306 @@
|
|
|
1
|
+
// Flash-Attention-Style Prefill Kernel (head_dim=256, f16 KV)
|
|
2
|
+
//
|
|
3
|
+
// Flash attention with KV-axis splitting to raise RDNA3 occupancy. The existing
|
|
4
|
+
// attention_head256_f16kv.wgsl dispatches (num_query_blocks × num_heads, 1, 1)
|
|
5
|
+
// workgroups, which is too few for RDNA3 at common prefill shapes (e.g. 16
|
|
6
|
+
// workgroups for prefill=64, 8 heads). This variant multiplies the workgroup
|
|
7
|
+
// count by num_kv_splits along the KV axis and produces partial (output, m, l)
|
|
8
|
+
// per split — a companion reduction kernel merges them.
|
|
9
|
+
//
|
|
10
|
+
// Algorithm: online-softmax flash attention per split.
|
|
11
|
+
//
|
|
12
|
+
// Pass 1 (this kernel): each workgroup processes one KV slice for its
|
|
13
|
+
// (query_block, head) pair; writes un-normalised acc + m + l to partial buffers.
|
|
14
|
+
// Pass 2 (flash_reduce): merges partials across splits for each (query, head).
|
|
15
|
+
|
|
16
|
+
enable f16;
|
|
17
|
+
|
|
18
|
+
const BLOCK_SIZE: u32 = 32u;
|
|
19
|
+
const WORKGROUP_SIZE: u32 = BLOCK_SIZE;
|
|
20
|
+
const HEAD_DIM: u32 = 256u;
|
|
21
|
+
const HEAD_DIM_VECS: u32 = 64u;
|
|
22
|
+
|
|
23
|
+
struct Uniforms {
|
|
24
|
+
num_heads: u32,
|
|
25
|
+
num_kv_heads: u32,
|
|
26
|
+
head_dim: u32,
|
|
27
|
+
seq_len: u32,
|
|
28
|
+
query_len: u32,
|
|
29
|
+
scale: f32,
|
|
30
|
+
is_causal: u32,
|
|
31
|
+
start_pos: u32,
|
|
32
|
+
attn_softcap: f32,
|
|
33
|
+
sliding_window: u32,
|
|
34
|
+
kv_len_source: u32,
|
|
35
|
+
kv_start: u32,
|
|
36
|
+
page_size: u32,
|
|
37
|
+
kv_layout: u32,
|
|
38
|
+
num_kv_splits: u32,
|
|
39
|
+
_pad: u32,
|
|
40
|
+
}
|
|
41
|
+
|
|
42
|
+
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
43
|
+
@group(0) @binding(1) var<storage, read> Q: array<f32>;
|
|
44
|
+
@group(0) @binding(2) var<storage, read> K: array<f16>;
|
|
45
|
+
@group(0) @binding(3) var<storage, read> V: array<f16>;
|
|
46
|
+
// Per-split partials: [num_query_blocks, num_heads, num_kv_splits, BLOCK_SIZE, HEAD_DIM]
|
|
47
|
+
// When num_kv_splits == 1 (single-split fast path), this buffer is bound to the
|
|
48
|
+
// FINAL output buffer and we write normalised output directly — skipping the
|
|
49
|
+
// reduce pass entirely.
|
|
50
|
+
@group(0) @binding(4) var<storage, read_write> partial_acc: array<f32>;
|
|
51
|
+
// Per-split stats: [num_query_blocks, num_heads, num_kv_splits, BLOCK_SIZE]
|
|
52
|
+
@group(0) @binding(5) var<storage, read_write> partial_m: array<f32>;
|
|
53
|
+
@group(0) @binding(6) var<storage, read_write> partial_l: array<f32>;
|
|
54
|
+
@group(0) @binding(7) var<storage, read> kv_len_buffer: array<u32>;
|
|
55
|
+
@group(0) @binding(8) var<storage, read> page_table: array<u32>;
|
|
56
|
+
|
|
57
|
+
// Separate shared tiles for K and V so they can be loaded in parallel within
|
|
58
|
+
// the same loop iteration. Saves a workgroupBarrier per KV block vs the
|
|
59
|
+
// shared-tile-reuse pattern of the non-flash head256 kernel.
|
|
60
|
+
var<workgroup> shared_K: array<vec4<f16>, BLOCK_SIZE * HEAD_DIM_VECS>;
|
|
61
|
+
var<workgroup> shared_V: array<vec4<f16>, BLOCK_SIZE * HEAD_DIM_VECS>;
|
|
62
|
+
|
|
63
|
+
fn zero_vec4_f16() -> vec4<f16> {
|
|
64
|
+
return vec4<f16>(f16(0.0), f16(0.0), f16(0.0), f16(0.0));
|
|
65
|
+
}
|
|
66
|
+
|
|
67
|
+
fn get_kv_head_idx(query_head_idx: u32) -> u32 {
|
|
68
|
+
let heads_per_kv = u.num_heads / u.num_kv_heads;
|
|
69
|
+
return query_head_idx / heads_per_kv;
|
|
70
|
+
}
|
|
71
|
+
|
|
72
|
+
fn get_kv_pos(key_pos: u32) -> u32 {
|
|
73
|
+
let abs_key = u.kv_start + key_pos;
|
|
74
|
+
if (u.kv_layout == 1u && u.sliding_window > 0u) {
|
|
75
|
+
return abs_key % u.sliding_window;
|
|
76
|
+
}
|
|
77
|
+
if (u.kv_layout == 2u) {
|
|
78
|
+
let page_idx = abs_key / u.page_size;
|
|
79
|
+
let in_page = abs_key - (page_idx * u.page_size);
|
|
80
|
+
let phys_page = page_table[page_idx];
|
|
81
|
+
return phys_page * u.page_size + in_page;
|
|
82
|
+
}
|
|
83
|
+
return abs_key;
|
|
84
|
+
}
|
|
85
|
+
|
|
86
|
+
fn get_kv_len() -> u32 {
|
|
87
|
+
if (u.kv_len_source == 0u) {
|
|
88
|
+
return u.seq_len;
|
|
89
|
+
}
|
|
90
|
+
return kv_len_buffer[0];
|
|
91
|
+
}
|
|
92
|
+
|
|
93
|
+
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
94
|
+
fn main(
|
|
95
|
+
@builtin(local_invocation_id) local_id: vec3<u32>,
|
|
96
|
+
@builtin(workgroup_id) wg_id: vec3<u32>
|
|
97
|
+
) {
|
|
98
|
+
if (u.head_dim != HEAD_DIM) {
|
|
99
|
+
return;
|
|
100
|
+
}
|
|
101
|
+
|
|
102
|
+
// Dispatch layout: linear = (query_block_idx * num_heads + head_idx) * num_kv_splits + kv_split_idx
|
|
103
|
+
let num_heads = u.num_heads;
|
|
104
|
+
let num_kv_splits = u.num_kv_splits;
|
|
105
|
+
let linear = wg_id.x;
|
|
106
|
+
let kv_split_idx = linear % num_kv_splits;
|
|
107
|
+
let linear_qh = linear / num_kv_splits;
|
|
108
|
+
let head_idx = linear_qh % num_heads;
|
|
109
|
+
let query_block_idx = linear_qh / num_heads;
|
|
110
|
+
let thread_idx = local_id.x;
|
|
111
|
+
|
|
112
|
+
let kv_head_idx = get_kv_head_idx(head_idx);
|
|
113
|
+
let seq_len = get_kv_len();
|
|
114
|
+
let query_len = u.query_len;
|
|
115
|
+
let scale = u.scale;
|
|
116
|
+
|
|
117
|
+
let query_pos = query_block_idx * BLOCK_SIZE + thread_idx;
|
|
118
|
+
let valid_query = query_pos < query_len;
|
|
119
|
+
let abs_query = query_pos + u.start_pos;
|
|
120
|
+
|
|
121
|
+
// KV slice bounds for this split
|
|
122
|
+
let kv_per_split = (seq_len + num_kv_splits - 1u) / num_kv_splits;
|
|
123
|
+
let kv_slice_start = kv_split_idx * kv_per_split;
|
|
124
|
+
let kv_slice_end = min(kv_slice_start + kv_per_split, seq_len);
|
|
125
|
+
|
|
126
|
+
var q_local: array<vec4<f32>, HEAD_DIM_VECS>;
|
|
127
|
+
var acc: array<vec4<f32>, HEAD_DIM_VECS>;
|
|
128
|
+
|
|
129
|
+
if (valid_query) {
|
|
130
|
+
let q_offset = query_pos * num_heads * HEAD_DIM + head_idx * HEAD_DIM;
|
|
131
|
+
for (var d4: u32 = 0u; d4 < HEAD_DIM_VECS; d4 = d4 + 1u) {
|
|
132
|
+
let base = q_offset + d4 * 4u;
|
|
133
|
+
q_local[d4] = vec4<f32>(Q[base], Q[base + 1u], Q[base + 2u], Q[base + 3u]);
|
|
134
|
+
acc[d4] = vec4<f32>(0.0);
|
|
135
|
+
}
|
|
136
|
+
}
|
|
137
|
+
|
|
138
|
+
var m_i: f32 = -3.402823e+38;
|
|
139
|
+
var l_i: f32 = 0.0;
|
|
140
|
+
|
|
141
|
+
// Clamp KV slice to causal/sliding constraints for this query
|
|
142
|
+
var min_key_pos: u32 = kv_slice_start;
|
|
143
|
+
var max_key_pos: u32 = kv_slice_end;
|
|
144
|
+
|
|
145
|
+
if (valid_query) {
|
|
146
|
+
if (u.is_causal != 0u) {
|
|
147
|
+
if (abs_query < u.kv_start) {
|
|
148
|
+
max_key_pos = min(max_key_pos, 0u);
|
|
149
|
+
} else {
|
|
150
|
+
let causal_limit = abs_query - u.kv_start + 1u;
|
|
151
|
+
max_key_pos = min(max_key_pos, causal_limit);
|
|
152
|
+
}
|
|
153
|
+
}
|
|
154
|
+
if (u.sliding_window > 0u && abs_query >= u.sliding_window) {
|
|
155
|
+
let min_abs_key = abs_query - u.sliding_window + 1u;
|
|
156
|
+
if (min_abs_key > u.kv_start) {
|
|
157
|
+
let sw_min = min_abs_key - u.kv_start;
|
|
158
|
+
if (sw_min > min_key_pos) {
|
|
159
|
+
min_key_pos = sw_min;
|
|
160
|
+
}
|
|
161
|
+
}
|
|
162
|
+
min_key_pos = min(min_key_pos, seq_len);
|
|
163
|
+
}
|
|
164
|
+
}
|
|
165
|
+
|
|
166
|
+
// Iterate KV blocks WITHIN our split range
|
|
167
|
+
let num_kv_blocks_in_slice = (kv_slice_end - kv_slice_start + BLOCK_SIZE - 1u) / BLOCK_SIZE;
|
|
168
|
+
|
|
169
|
+
for (var kv_block: u32 = 0u; kv_block < num_kv_blocks_in_slice; kv_block = kv_block + 1u) {
|
|
170
|
+
let kv_block_start = kv_slice_start + kv_block * BLOCK_SIZE;
|
|
171
|
+
|
|
172
|
+
var scores: array<f32, BLOCK_SIZE>;
|
|
173
|
+
var key_active: array<u32, BLOCK_SIZE>;
|
|
174
|
+
var probs: array<f32, BLOCK_SIZE>;
|
|
175
|
+
for (var k_init: u32 = 0u; k_init < BLOCK_SIZE; k_init = k_init + 1u) {
|
|
176
|
+
scores[k_init] = 0.0;
|
|
177
|
+
key_active[k_init] = 0u;
|
|
178
|
+
probs[k_init] = 0.0;
|
|
179
|
+
}
|
|
180
|
+
|
|
181
|
+
if (valid_query) {
|
|
182
|
+
for (var k_mask: u32 = 0u; k_mask < BLOCK_SIZE; k_mask = k_mask + 1u) {
|
|
183
|
+
let key_pos = kv_block_start + k_mask;
|
|
184
|
+
if (key_pos < kv_slice_end && key_pos >= min_key_pos && key_pos < max_key_pos) {
|
|
185
|
+
key_active[k_mask] = 1u;
|
|
186
|
+
}
|
|
187
|
+
}
|
|
188
|
+
}
|
|
189
|
+
|
|
190
|
+
// Load K and V in parallel into separate shared tiles. A single barrier
|
|
191
|
+
// synchronises both loads before we score K and accumulate V.
|
|
192
|
+
let key_pos_load = kv_block_start + thread_idx;
|
|
193
|
+
let shared_row = thread_idx * HEAD_DIM_VECS;
|
|
194
|
+
if (key_pos_load < kv_slice_end) {
|
|
195
|
+
let k_idx = get_kv_pos(key_pos_load);
|
|
196
|
+
let k_offset = k_idx * u.num_kv_heads * HEAD_DIM + kv_head_idx * HEAD_DIM;
|
|
197
|
+
let v_idx = k_idx;
|
|
198
|
+
let v_offset = v_idx * u.num_kv_heads * HEAD_DIM + kv_head_idx * HEAD_DIM;
|
|
199
|
+
for (var d4: u32 = 0u; d4 < HEAD_DIM_VECS; d4 = d4 + 1u) {
|
|
200
|
+
let kb = k_offset + d4 * 4u;
|
|
201
|
+
let vb = v_offset + d4 * 4u;
|
|
202
|
+
shared_K[shared_row + d4] = vec4<f16>(
|
|
203
|
+
K[kb], K[kb + 1u], K[kb + 2u], K[kb + 3u]
|
|
204
|
+
);
|
|
205
|
+
shared_V[shared_row + d4] = vec4<f16>(
|
|
206
|
+
V[vb], V[vb + 1u], V[vb + 2u], V[vb + 3u]
|
|
207
|
+
);
|
|
208
|
+
}
|
|
209
|
+
} else {
|
|
210
|
+
for (var d4: u32 = 0u; d4 < HEAD_DIM_VECS; d4 = d4 + 1u) {
|
|
211
|
+
shared_K[shared_row + d4] = zero_vec4_f16();
|
|
212
|
+
shared_V[shared_row + d4] = zero_vec4_f16();
|
|
213
|
+
}
|
|
214
|
+
}
|
|
215
|
+
|
|
216
|
+
workgroupBarrier();
|
|
217
|
+
|
|
218
|
+
var m_new: f32 = m_i;
|
|
219
|
+
if (valid_query) {
|
|
220
|
+
var block_max: f32 = -3.402823e+38;
|
|
221
|
+
for (var k: u32 = 0u; k < BLOCK_SIZE; k = k + 1u) {
|
|
222
|
+
if (key_active[k] == 0u) { continue; }
|
|
223
|
+
let key_row = k * HEAD_DIM_VECS;
|
|
224
|
+
var dot_partial: f32 = 0.0;
|
|
225
|
+
for (var d4: u32 = 0u; d4 < HEAD_DIM_VECS; d4 = d4 + 1u) {
|
|
226
|
+
dot_partial = dot_partial + dot(
|
|
227
|
+
q_local[d4], vec4<f32>(shared_K[key_row + d4])
|
|
228
|
+
);
|
|
229
|
+
}
|
|
230
|
+
var s = dot_partial * scale;
|
|
231
|
+
if (u.attn_softcap > 0.0) {
|
|
232
|
+
s = tanh(s / u.attn_softcap) * u.attn_softcap;
|
|
233
|
+
}
|
|
234
|
+
scores[k] = s;
|
|
235
|
+
block_max = max(block_max, s);
|
|
236
|
+
}
|
|
237
|
+
|
|
238
|
+
m_new = max(m_i, block_max);
|
|
239
|
+
let correction = exp(m_i - m_new);
|
|
240
|
+
l_i = l_i * correction;
|
|
241
|
+
for (var d4: u32 = 0u; d4 < HEAD_DIM_VECS; d4 = d4 + 1u) {
|
|
242
|
+
acc[d4] = acc[d4] * correction;
|
|
243
|
+
}
|
|
244
|
+
|
|
245
|
+
for (var k: u32 = 0u; k < BLOCK_SIZE; k = k + 1u) {
|
|
246
|
+
if (key_active[k] == 0u) { continue; }
|
|
247
|
+
let p = exp(scores[k] - m_new);
|
|
248
|
+
probs[k] = p;
|
|
249
|
+
l_i = l_i + p;
|
|
250
|
+
}
|
|
251
|
+
|
|
252
|
+
// V is already in shared_V — apply immediately (no reload, no barrier).
|
|
253
|
+
for (var k: u32 = 0u; k < BLOCK_SIZE; k = k + 1u) {
|
|
254
|
+
let p = probs[k];
|
|
255
|
+
if (p == 0.0) { continue; }
|
|
256
|
+
let value_row = k * HEAD_DIM_VECS;
|
|
257
|
+
for (var d4: u32 = 0u; d4 < HEAD_DIM_VECS; d4 = d4 + 1u) {
|
|
258
|
+
acc[d4] = acc[d4] + p * vec4<f32>(shared_V[value_row + d4]);
|
|
259
|
+
}
|
|
260
|
+
}
|
|
261
|
+
m_i = m_new;
|
|
262
|
+
}
|
|
263
|
+
|
|
264
|
+
workgroupBarrier();
|
|
265
|
+
}
|
|
266
|
+
|
|
267
|
+
if (valid_query) {
|
|
268
|
+
let qh_idx = query_block_idx * num_heads + head_idx;
|
|
269
|
+
if (num_kv_splits == 1u) {
|
|
270
|
+
// Single-split fast path: skip the reduce pass entirely. Write the
|
|
271
|
+
// FINAL normalised output directly to the output buffer (which is
|
|
272
|
+
// bound in slot 4 under the partial_acc name). Layout matches
|
|
273
|
+
// [query_len, num_heads, head_dim].
|
|
274
|
+
let inv_l_i = select(0.0, 1.0 / l_i, l_i > 0.0);
|
|
275
|
+
let out_offset = query_pos * num_heads * HEAD_DIM + head_idx * HEAD_DIM;
|
|
276
|
+
for (var d4: u32 = 0u; d4 < HEAD_DIM_VECS; d4 = d4 + 1u) {
|
|
277
|
+
let base = out_offset + d4 * 4u;
|
|
278
|
+
let normed = acc[d4] * inv_l_i;
|
|
279
|
+
partial_acc[base] = normed.x;
|
|
280
|
+
partial_acc[base + 1u] = normed.y;
|
|
281
|
+
partial_acc[base + 2u] = normed.z;
|
|
282
|
+
partial_acc[base + 3u] = normed.w;
|
|
283
|
+
}
|
|
284
|
+
// m/l buffers are still bound but unused on the single-split path.
|
|
285
|
+
// Leave them untouched to avoid spurious writes.
|
|
286
|
+
} else {
|
|
287
|
+
// Multi-split: write UN-NORMALISED partial acc and per-query m/l for
|
|
288
|
+
// this split. Layout:
|
|
289
|
+
// partial_acc[(qb*num_heads + h) * num_kv_splits * BLOCK_SIZE * HEAD_DIM
|
|
290
|
+
// + kv_split * BLOCK_SIZE * HEAD_DIM
|
|
291
|
+
// + thread_idx * HEAD_DIM + d]
|
|
292
|
+
let partial_base = (qh_idx * num_kv_splits + kv_split_idx) * BLOCK_SIZE * HEAD_DIM
|
|
293
|
+
+ thread_idx * HEAD_DIM;
|
|
294
|
+
for (var d4: u32 = 0u; d4 < HEAD_DIM_VECS; d4 = d4 + 1u) {
|
|
295
|
+
let base = partial_base + d4 * 4u;
|
|
296
|
+
partial_acc[base] = acc[d4].x;
|
|
297
|
+
partial_acc[base + 1u] = acc[d4].y;
|
|
298
|
+
partial_acc[base + 2u] = acc[d4].z;
|
|
299
|
+
partial_acc[base + 3u] = acc[d4].w;
|
|
300
|
+
}
|
|
301
|
+
let stats_idx = (qh_idx * num_kv_splits + kv_split_idx) * BLOCK_SIZE + thread_idx;
|
|
302
|
+
partial_m[stats_idx] = m_i;
|
|
303
|
+
partial_l[stats_idx] = l_i;
|
|
304
|
+
}
|
|
305
|
+
}
|
|
306
|
+
}
|
|
@@ -0,0 +1,101 @@
|
|
|
1
|
+
// Flash-Attention Reduction Kernel
|
|
2
|
+
//
|
|
3
|
+
// Merges per-split (partial_acc, m, l) produced by
|
|
4
|
+
// attention_prefill_flash_head256_f16kv.wgsl into the final attention output.
|
|
5
|
+
//
|
|
6
|
+
// Online softmax merge across splits: for each (query, head),
|
|
7
|
+
// m_global = max(m_s)
|
|
8
|
+
// l_global = sum(exp(m_s - m_global) * l_s)
|
|
9
|
+
// output[d] = sum(exp(m_s - m_global) * acc_s[d]) / l_global
|
|
10
|
+
//
|
|
11
|
+
// Dispatch: (ceil(query_len * num_heads / WORKGROUP_SIZE), HEAD_DIM_VECS, 1)
|
|
12
|
+
// Each thread owns one (query_head_flat, d4) output lane.
|
|
13
|
+
|
|
14
|
+
const HEAD_DIM: u32 = 256u;
|
|
15
|
+
const HEAD_DIM_VECS: u32 = 64u;
|
|
16
|
+
const BLOCK_SIZE: u32 = 32u;
|
|
17
|
+
const WORKGROUP_SIZE: u32 = 64u;
|
|
18
|
+
const MAX_KV_SPLITS: u32 = 16u;
|
|
19
|
+
|
|
20
|
+
struct ReduceUniforms {
|
|
21
|
+
num_heads: u32,
|
|
22
|
+
query_len: u32,
|
|
23
|
+
num_kv_splits: u32,
|
|
24
|
+
_pad: u32,
|
|
25
|
+
}
|
|
26
|
+
|
|
27
|
+
@group(0) @binding(0) var<uniform> u: ReduceUniforms;
|
|
28
|
+
@group(0) @binding(1) var<storage, read> partial_acc: array<f32>;
|
|
29
|
+
@group(0) @binding(2) var<storage, read> partial_m: array<f32>;
|
|
30
|
+
@group(0) @binding(3) var<storage, read> partial_l: array<f32>;
|
|
31
|
+
@group(0) @binding(4) var<storage, read_write> output: array<f32>;
|
|
32
|
+
|
|
33
|
+
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
34
|
+
fn main(
|
|
35
|
+
@builtin(global_invocation_id) global_id: vec3<u32>
|
|
36
|
+
) {
|
|
37
|
+
let qh_flat = global_id.x;
|
|
38
|
+
let d4 = global_id.y;
|
|
39
|
+
|
|
40
|
+
let num_heads = u.num_heads;
|
|
41
|
+
let query_len = u.query_len;
|
|
42
|
+
let num_kv_splits = u.num_kv_splits;
|
|
43
|
+
|
|
44
|
+
let total_qh = query_len * num_heads;
|
|
45
|
+
if (qh_flat >= total_qh || d4 >= HEAD_DIM_VECS) {
|
|
46
|
+
return;
|
|
47
|
+
}
|
|
48
|
+
|
|
49
|
+
let query_pos = qh_flat / num_heads;
|
|
50
|
+
let head_idx = qh_flat % num_heads;
|
|
51
|
+
let query_block_idx = query_pos / BLOCK_SIZE;
|
|
52
|
+
let thread_within_block = query_pos - query_block_idx * BLOCK_SIZE;
|
|
53
|
+
let qh_idx = query_block_idx * num_heads + head_idx;
|
|
54
|
+
|
|
55
|
+
// Gather m_s and l_s across splits to find global m and global l.
|
|
56
|
+
var m_global: f32 = -3.402823e+38;
|
|
57
|
+
// Safe cap on splits so compilers accept the bounded loop.
|
|
58
|
+
let splits = min(num_kv_splits, MAX_KV_SPLITS);
|
|
59
|
+
for (var s: u32 = 0u; s < splits; s = s + 1u) {
|
|
60
|
+
let stats_idx = (qh_idx * num_kv_splits + s) * BLOCK_SIZE + thread_within_block;
|
|
61
|
+
let m_s = partial_m[stats_idx];
|
|
62
|
+
if (m_s > m_global) {
|
|
63
|
+
m_global = m_s;
|
|
64
|
+
}
|
|
65
|
+
}
|
|
66
|
+
|
|
67
|
+
// Compute global l as weighted sum of per-split l with exp correction.
|
|
68
|
+
var l_global: f32 = 0.0;
|
|
69
|
+
for (var s: u32 = 0u; s < splits; s = s + 1u) {
|
|
70
|
+
let stats_idx = (qh_idx * num_kv_splits + s) * BLOCK_SIZE + thread_within_block;
|
|
71
|
+
let m_s = partial_m[stats_idx];
|
|
72
|
+
let l_s = partial_l[stats_idx];
|
|
73
|
+
l_global = l_global + l_s * exp(m_s - m_global);
|
|
74
|
+
}
|
|
75
|
+
|
|
76
|
+
// Merge acc across splits for this d4.
|
|
77
|
+
var merged = vec4<f32>(0.0);
|
|
78
|
+
for (var s: u32 = 0u; s < splits; s = s + 1u) {
|
|
79
|
+
let stats_idx = (qh_idx * num_kv_splits + s) * BLOCK_SIZE + thread_within_block;
|
|
80
|
+
let m_s = partial_m[stats_idx];
|
|
81
|
+
let correction = exp(m_s - m_global);
|
|
82
|
+
let partial_base = (qh_idx * num_kv_splits + s) * BLOCK_SIZE * HEAD_DIM
|
|
83
|
+
+ thread_within_block * HEAD_DIM
|
|
84
|
+
+ d4 * 4u;
|
|
85
|
+
let partial_vec = vec4<f32>(
|
|
86
|
+
partial_acc[partial_base],
|
|
87
|
+
partial_acc[partial_base + 1u],
|
|
88
|
+
partial_acc[partial_base + 2u],
|
|
89
|
+
partial_acc[partial_base + 3u]
|
|
90
|
+
);
|
|
91
|
+
merged = merged + correction * partial_vec;
|
|
92
|
+
}
|
|
93
|
+
|
|
94
|
+
let inv_l = select(0.0, 1.0 / l_global, l_global > 0.0);
|
|
95
|
+
let out_offset = query_pos * num_heads * HEAD_DIM + head_idx * HEAD_DIM + d4 * 4u;
|
|
96
|
+
let final_vec = merged * inv_l;
|
|
97
|
+
output[out_offset] = final_vec.x;
|
|
98
|
+
output[out_offset + 1u] = final_vec.y;
|
|
99
|
+
output[out_offset + 2u] = final_vec.z;
|
|
100
|
+
output[out_offset + 3u] = final_vec.w;
|
|
101
|
+
}
|
|
@@ -1,9 +1,9 @@
|
|
|
1
1
|
// AUTO-GENERATED from src/gpu/kernels/attention_small.wgsl.
|
|
2
|
-
// Edit the source kernel and src/gpu/kernels/codegen/wgsl-variants.js, then run `npm run kernels:
|
|
2
|
+
// Edit the source kernel and src/gpu/kernels/codegen/wgsl-variants.js, then run `npm run kernels:codegen:sync`.
|
|
3
3
|
// Fused Multi-Head Attention Kernel (small tiles, f16 QKV + f16 output)
|
|
4
4
|
//
|
|
5
5
|
// Same algorithm as attention_small.wgsl but Q/K/V are f16 and output is f16.
|
|
6
|
-
// Shared tiles
|
|
6
|
+
// Shared tiles and attention arithmetic use f16 for the all-f16 lane.
|
|
7
7
|
|
|
8
8
|
enable f16;
|
|
9
9
|
|
|
@@ -11,6 +11,7 @@ const BLOCK_SIZE: u32 = 32u;
|
|
|
11
11
|
const WORKGROUP_SIZE: u32 = BLOCK_SIZE;
|
|
12
12
|
const HEAD_TILE: u32 = 32u;
|
|
13
13
|
const MAX_HEAD_DIM: u32 = 256u;
|
|
14
|
+
const NEG_INF = f16(-65504.0);
|
|
14
15
|
|
|
15
16
|
struct Uniforms {
|
|
16
17
|
num_heads: u32,
|
|
@@ -95,24 +96,25 @@ fn main(
|
|
|
95
96
|
}
|
|
96
97
|
let seq_len = get_kv_len();
|
|
97
98
|
let query_len = u.query_len;
|
|
98
|
-
let scale = u.scale;
|
|
99
|
+
let scale = f16(u.scale);
|
|
100
|
+
let softcap = f16(u.attn_softcap);
|
|
99
101
|
|
|
100
102
|
let query_pos = query_block_idx * BLOCK_SIZE + thread_idx;
|
|
101
103
|
let valid_query = query_pos < query_len;
|
|
102
104
|
|
|
103
|
-
var q_local: array<
|
|
104
|
-
var acc: array<
|
|
105
|
+
var q_local: array<f16, MAX_HEAD_DIM>;
|
|
106
|
+
var acc: array<f16, MAX_HEAD_DIM>;
|
|
105
107
|
|
|
106
108
|
if (valid_query) {
|
|
107
109
|
let q_offset = query_pos * num_heads * head_dim + head_idx * head_dim;
|
|
108
110
|
for (var d: u32 = 0u; d < head_dim; d = d + 1u) {
|
|
109
|
-
q_local[d] =
|
|
110
|
-
acc[d] = 0.0;
|
|
111
|
+
q_local[d] = Q[q_offset + d];
|
|
112
|
+
acc[d] = f16(0.0);
|
|
111
113
|
}
|
|
112
114
|
}
|
|
113
115
|
|
|
114
|
-
var m_i:
|
|
115
|
-
var l_i:
|
|
116
|
+
var m_i: f16 = NEG_INF;
|
|
117
|
+
var l_i: f16 = f16(0.0);
|
|
116
118
|
|
|
117
119
|
let num_kv_blocks = (seq_len + BLOCK_SIZE - 1u) / BLOCK_SIZE;
|
|
118
120
|
let num_head_tiles = (head_dim + HEAD_TILE - 1u) / HEAD_TILE;
|
|
@@ -120,9 +122,9 @@ fn main(
|
|
|
120
122
|
for (var kv_block: u32 = 0u; kv_block < num_kv_blocks; kv_block = kv_block + 1u) {
|
|
121
123
|
let kv_block_start = kv_block * BLOCK_SIZE;
|
|
122
124
|
|
|
123
|
-
var scores: array<
|
|
125
|
+
var scores: array<f16, BLOCK_SIZE>;
|
|
124
126
|
for (var k_init: u32 = 0u; k_init < BLOCK_SIZE; k_init = k_init + 1u) {
|
|
125
|
-
scores[k_init] = 0.0;
|
|
127
|
+
scores[k_init] = f16(0.0);
|
|
126
128
|
}
|
|
127
129
|
|
|
128
130
|
for (var ht: u32 = 0u; ht < num_head_tiles; ht = ht + 1u) {
|
|
@@ -150,9 +152,9 @@ fn main(
|
|
|
150
152
|
if (key_pos >= seq_len) { continue; }
|
|
151
153
|
if (is_masked(query_pos, key_pos)) { continue; }
|
|
152
154
|
|
|
153
|
-
var dot_partial:
|
|
155
|
+
var dot_partial: f16 = f16(0.0);
|
|
154
156
|
for (var td: u32 = 0u; td < tile_len; td = td + 1u) {
|
|
155
|
-
dot_partial = dot_partial + q_local[d0 + td] *
|
|
157
|
+
dot_partial = dot_partial + q_local[d0 + td] * shared_K[k * HEAD_TILE + td];
|
|
156
158
|
}
|
|
157
159
|
scores[k] = scores[k] + dot_partial;
|
|
158
160
|
}
|
|
@@ -161,17 +163,17 @@ fn main(
|
|
|
161
163
|
workgroupBarrier();
|
|
162
164
|
}
|
|
163
165
|
|
|
164
|
-
var m_new:
|
|
166
|
+
var m_new: f16 = m_i;
|
|
165
167
|
if (valid_query) {
|
|
166
|
-
var block_max:
|
|
168
|
+
var block_max: f16 = NEG_INF;
|
|
167
169
|
for (var k: u32 = 0u; k < BLOCK_SIZE; k = k + 1u) {
|
|
168
170
|
let key_pos = kv_block_start + k;
|
|
169
171
|
if (key_pos >= seq_len) { continue; }
|
|
170
172
|
if (is_masked(query_pos, key_pos)) { continue; }
|
|
171
173
|
|
|
172
174
|
var s = scores[k] * scale;
|
|
173
|
-
if (
|
|
174
|
-
s = tanh(s /
|
|
175
|
+
if (softcap > f16(0.0)) {
|
|
176
|
+
s = tanh(s / softcap) * softcap;
|
|
175
177
|
}
|
|
176
178
|
scores[k] = s;
|
|
177
179
|
block_max = max(block_max, s);
|
|
@@ -217,7 +219,7 @@ fn main(
|
|
|
217
219
|
}
|
|
218
220
|
|
|
219
221
|
for (var td: u32 = 0u; td < tile_len; td = td + 1u) {
|
|
220
|
-
acc[d0 + td] = acc[d0 + td] + p *
|
|
222
|
+
acc[d0 + td] = acc[d0 + td] + p * shared_V[k * HEAD_TILE + td];
|
|
221
223
|
}
|
|
222
224
|
}
|
|
223
225
|
}
|
|
@@ -232,9 +234,9 @@ fn main(
|
|
|
232
234
|
|
|
233
235
|
if (valid_query) {
|
|
234
236
|
let out_offset = query_pos * num_heads * head_dim + head_idx * head_dim;
|
|
235
|
-
let inv_l_i = select(0.0, 1.0 / l_i, l_i > 0.0);
|
|
237
|
+
let inv_l_i = select(f16(0.0), f16(1.0) / l_i, l_i > f16(0.0));
|
|
236
238
|
for (var d: u32 = 0u; d < head_dim; d = d + 1u) {
|
|
237
|
-
output[out_offset + d] =
|
|
239
|
+
output[out_offset + d] = acc[d] * inv_l_i;
|
|
238
240
|
}
|
|
239
241
|
}
|
|
240
242
|
}
|
|
@@ -1,5 +1,5 @@
|
|
|
1
1
|
// AUTO-GENERATED from src/gpu/kernels/attention_small.wgsl.
|
|
2
|
-
// Edit the source kernel and src/gpu/kernels/codegen/wgsl-variants.js, then run `npm run kernels:
|
|
2
|
+
// Edit the source kernel and src/gpu/kernels/codegen/wgsl-variants.js, then run `npm run kernels:codegen:sync`.
|
|
3
3
|
// Fused Multi-Head Attention Kernel (small tiles, f16 KV)
|
|
4
4
|
//
|
|
5
5
|
// Same algorithm as attention_small.wgsl but K/V are stored as f16.
|
|
@@ -1,10 +1,10 @@
|
|
|
1
1
|
// AUTO-GENERATED from src/gpu/kernels/attention_streaming.wgsl.
|
|
2
|
-
// Edit the source kernel and src/gpu/kernels/codegen/wgsl-variants.js, then run `npm run kernels:
|
|
3
|
-
// Streaming Multi-Head Attention Kernel (
|
|
2
|
+
// Edit the source kernel and src/gpu/kernels/codegen/wgsl-variants.js, then run `npm run kernels:codegen:sync`.
|
|
3
|
+
// Streaming Multi-Head Attention Kernel (no workgroup storage)
|
|
4
4
|
//
|
|
5
|
-
//
|
|
6
|
-
|
|
7
|
-
|
|
5
|
+
// Fallback variant for devices with extremely small shared memory or
|
|
6
|
+
// models with head_dim beyond tiled support. Uses two-pass softmax and
|
|
7
|
+
// reads K/V directly from storage. Slower but compatible.
|
|
8
8
|
|
|
9
9
|
override WORKGROUP_SIZE: u32 = 1u;
|
|
10
10
|
const MAX_HEAD_DIM: u32 = 512u;
|
|
@@ -17,14 +17,18 @@ struct Uniforms {
|
|
|
17
17
|
query_len: u32,
|
|
18
18
|
scale: f32,
|
|
19
19
|
is_causal: u32,
|
|
20
|
-
start_pos: u32,
|
|
21
|
-
attn_softcap: f32,
|
|
22
|
-
sliding_window: u32,
|
|
23
|
-
kv_len_source: u32,
|
|
20
|
+
start_pos: u32, // Absolute position offset for causal masking
|
|
21
|
+
attn_softcap: f32, // Gemma 2: 50.0, 0 = disabled
|
|
22
|
+
sliding_window: u32, // Sliding window size (0 = disabled, >0 = window size)
|
|
23
|
+
kv_len_source: u32, // 0 = use uniform seq_len, 1 = use buffer
|
|
24
24
|
kv_start: u32,
|
|
25
25
|
page_size: u32,
|
|
26
26
|
kv_layout: u32,
|
|
27
|
-
|
|
27
|
+
bidirectional_span_start: u32,
|
|
28
|
+
bidirectional_span_length: u32,
|
|
29
|
+
_pad0: u32,
|
|
30
|
+
_pad1: u32,
|
|
31
|
+
_pad2: u32,
|
|
28
32
|
}
|
|
29
33
|
|
|
30
34
|
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
@@ -40,10 +44,27 @@ fn get_kv_head_idx(query_head_idx: u32) -> u32 {
|
|
|
40
44
|
return query_head_idx / heads_per_kv;
|
|
41
45
|
}
|
|
42
46
|
|
|
47
|
+
fn is_bidirectional_span_visible(abs_query: u32, abs_key: u32) -> bool {
|
|
48
|
+
if (u.bidirectional_span_length == 0u) {
|
|
49
|
+
return false;
|
|
50
|
+
}
|
|
51
|
+
let span_start = u.bidirectional_span_start;
|
|
52
|
+
let span_end = span_start + u.bidirectional_span_length;
|
|
53
|
+
return abs_query >= span_start
|
|
54
|
+
&& abs_query < span_end
|
|
55
|
+
&& abs_key >= span_start
|
|
56
|
+
&& abs_key < span_end;
|
|
57
|
+
}
|
|
58
|
+
|
|
43
59
|
fn is_masked(query_pos: u32, key_pos: u32) -> bool {
|
|
44
60
|
let abs_query = query_pos + u.start_pos;
|
|
45
61
|
let abs_key = u.kv_start + key_pos;
|
|
46
|
-
|
|
62
|
+
// Causal mask
|
|
63
|
+
if (u.is_causal != 0u && abs_key > abs_query) {
|
|
64
|
+
if (is_bidirectional_span_visible(abs_query, abs_key)) { return false; }
|
|
65
|
+
return true;
|
|
66
|
+
}
|
|
67
|
+
// Sliding window mask
|
|
47
68
|
if (u.sliding_window > 0u && abs_query >= u.sliding_window) {
|
|
48
69
|
if (abs_key < abs_query - u.sliding_window + 1u) { return true; }
|
|
49
70
|
}
|
|
@@ -99,9 +120,10 @@ fn main(@builtin(workgroup_id) wg_id: vec3<u32>) {
|
|
|
99
120
|
let k_offset = k_idx * u.num_kv_heads * head_dim + kv_head_idx * head_dim;
|
|
100
121
|
var dot: f32 = 0.0;
|
|
101
122
|
for (var d: u32 = 0u; d < head_dim; d = d + 1u) {
|
|
102
|
-
dot = dot + q_local[d] *
|
|
123
|
+
dot = dot + q_local[d] * K[k_offset + d];
|
|
103
124
|
}
|
|
104
125
|
dot = dot * scale;
|
|
126
|
+
// Gemma 2 attention softcapping
|
|
105
127
|
if (u.attn_softcap > 0.0) {
|
|
106
128
|
dot = tanh(dot / u.attn_softcap) * u.attn_softcap;
|
|
107
129
|
}
|
|
@@ -121,27 +143,28 @@ fn main(@builtin(workgroup_id) wg_id: vec3<u32>) {
|
|
|
121
143
|
let v_offset = k_idx * u.num_kv_heads * head_dim + kv_head_idx * head_dim;
|
|
122
144
|
var dot: f32 = 0.0;
|
|
123
145
|
for (var d: u32 = 0u; d < head_dim; d = d + 1u) {
|
|
124
|
-
dot = dot + q_local[d] *
|
|
146
|
+
dot = dot + q_local[d] * K[k_offset + d];
|
|
125
147
|
}
|
|
126
148
|
dot = dot * scale;
|
|
149
|
+
// Gemma 2 attention softcapping
|
|
127
150
|
if (u.attn_softcap > 0.0) {
|
|
128
151
|
dot = tanh(dot / u.attn_softcap) * u.attn_softcap;
|
|
129
152
|
}
|
|
130
153
|
let w = exp(dot - max_score);
|
|
131
154
|
sum_exp = sum_exp + w;
|
|
132
155
|
for (var d: u32 = 0u; d < head_dim; d = d + 1u) {
|
|
133
|
-
acc[d] = acc[d] + w *
|
|
156
|
+
acc[d] = acc[d] + w * V[v_offset + d];
|
|
134
157
|
}
|
|
135
158
|
}
|
|
136
159
|
|
|
137
160
|
let out_offset = query_pos * num_heads * head_dim + head_idx * head_dim;
|
|
138
161
|
if (sum_exp <= 0.0) {
|
|
139
162
|
for (var d: u32 = 0u; d < head_dim; d = d + 1u) {
|
|
140
|
-
output[out_offset + d] =
|
|
163
|
+
output[out_offset + d] = 0.0;
|
|
141
164
|
}
|
|
142
165
|
return;
|
|
143
166
|
}
|
|
144
167
|
for (var d: u32 = 0u; d < head_dim; d = d + 1u) {
|
|
145
|
-
output[out_offset + d] =
|
|
168
|
+
output[out_offset + d] = acc[d] / sum_exp;
|
|
146
169
|
}
|
|
147
170
|
}
|