doppler-gpu 0.2.2 → 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 +20 -8
- package/assets/doppler.svg +1601 -0
- package/package.json +82 -33
- 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 +20 -4
- 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 +79 -5
- 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 +368 -62
- 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 +615 -368
- 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 +32 -0
- 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/benchmark.schema.d.ts +2 -0
- package/src/config/schema/benchmark.schema.js +1 -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 +21 -3
- package/src/config/schema/doppler.schema.js +116 -9
- 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 +54 -1
- package/src/config/schema/index.js +15 -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 +73 -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 +81 -4
- 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/source-transform-contract.js +390 -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 +115 -1
- package/src/formats/rdrr/validation.js +245 -0
- package/src/formats/tflite/types.d.ts +8 -31
- package/src/formats/tflite/types.js +50 -5
- package/src/gpu/command-recorder.d.ts +9 -21
- package/src/gpu/command-recorder.js +41 -2
- 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.d.ts +22 -0
- package/src/gpu/kernels/activation-static-qdq.js +69 -0
- package/src/gpu/kernels/activation_static_qdq.wgsl +25 -0
- 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 +13 -5
- package/src/gpu/kernels/index.js +9 -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 +13 -0
- package/src/inference/browser-harness-text-helpers.js +267 -18
- package/src/inference/browser-harness.d.ts +0 -29
- package/src/inference/browser-harness.js +271 -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 +21 -2
- 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/attention/types.js +9 -4
- 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 +23 -27
- package/src/inference/pipelines/text/embed.js +10 -43
- 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 +6 -2
- package/src/inference/pipelines/text/generator-helpers.js +20 -2
- 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 +23 -9
- package/src/inference/pipelines/text/generator.d.ts +0 -6
- package/src/inference/pipelines/text/generator.js +254 -530
- 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 +8 -2
- package/src/inference/pipelines/text/init.js +190 -710
- 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.d.ts +5 -4
- package/src/inference/pipelines/text/logits/gpu.js +148 -11
- 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 +347 -59
- package/src/inference/pipelines/text/probes.d.ts +3 -6
- package/src/inference/pipelines/text/probes.js +198 -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 +12 -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 +233 -29
- 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 +30 -5
- package/src/loader/loader-types.d.ts +4 -0
- package/src/loader/manifest-config.d.ts +2 -7
- package/src/loader/manifest-config.js +25 -1
- package/src/loader/memory-monitor.d.ts +17 -5
- package/src/loader/per-layer-input-loader.js +114 -1
- 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.d.ts +15 -0
- package/src/loader/tensors/source-transform.js +317 -34
- package/src/loader/tensors/tensor-loader.d.ts +2 -9
- package/src/loader/tensors/tensor-loader.js +84 -1
- package/src/loader/tensors/tensor-reader.js +28 -2
- 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 +12 -3
- package/src/storage/artifact-storage-context.js +389 -98
- 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 +22 -0
- package/src/storage/quickstart-downloader.js +44 -56
- 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 +373 -21
- 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 +257 -150
- 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 +27 -0
- package/src/tooling/source-runtime-bundle.js +56 -23
- 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 +4 -8
- 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 +1,25 @@
|
|
|
1
|
+
override WORKGROUP_SIZE: u32 = 256u;
|
|
2
|
+
|
|
3
|
+
struct Uniforms {
|
|
4
|
+
size: u32,
|
|
5
|
+
_pad0: u32,
|
|
6
|
+
scale: f32,
|
|
7
|
+
invScale: f32,
|
|
8
|
+
qmin: f32,
|
|
9
|
+
qmax: f32,
|
|
10
|
+
}
|
|
11
|
+
|
|
12
|
+
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
13
|
+
@group(0) @binding(1) var<storage, read> input: array<f32>;
|
|
14
|
+
@group(0) @binding(2) var<storage, read_write> output: array<f32>;
|
|
15
|
+
|
|
16
|
+
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
17
|
+
fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
|
|
18
|
+
let idx = gid.x;
|
|
19
|
+
if (idx >= u.size) {
|
|
20
|
+
return;
|
|
21
|
+
}
|
|
22
|
+
|
|
23
|
+
let quantized = clamp(round(input[idx] * u.invScale), u.qmin, u.qmax);
|
|
24
|
+
output[idx] = quantized * u.scale;
|
|
25
|
+
}
|
|
@@ -96,25 +96,6 @@ export interface BDPAAttentionOptions extends OutputBufferOptions {
|
|
|
96
96
|
|
|
97
97
|
export type AttentionTier = 'subgroup' | 'tiled_large' | 'tiled_small' | 'streaming';
|
|
98
98
|
|
|
99
|
-
/** Context for attention tier selection rules. */
|
|
100
|
-
export interface AttentionTierContext {
|
|
101
|
-
canSubgroup: boolean;
|
|
102
|
-
canLarge: boolean;
|
|
103
|
-
canSmall: boolean;
|
|
104
|
-
isDecode: boolean;
|
|
105
|
-
}
|
|
106
|
-
|
|
107
|
-
/** Context for attention variant selection rules. */
|
|
108
|
-
export interface AttentionVariantContext {
|
|
109
|
-
tier: AttentionTier;
|
|
110
|
-
useF16KV: boolean;
|
|
111
|
-
canUseChunked: boolean;
|
|
112
|
-
canUseDecodeSubgroup: boolean;
|
|
113
|
-
canUseDecodeOptimized?: boolean;
|
|
114
|
-
isPaged?: boolean;
|
|
115
|
-
isDecode?: boolean;
|
|
116
|
-
}
|
|
117
|
-
|
|
118
99
|
/**
|
|
119
100
|
* Run attention operation
|
|
120
101
|
*/
|
|
@@ -236,3 +217,31 @@ export declare function resolveAttentionPlanForTest(
|
|
|
236
217
|
useF16KV: boolean;
|
|
237
218
|
isDecode: boolean;
|
|
238
219
|
};
|
|
220
|
+
|
|
221
|
+
/**
|
|
222
|
+
* Run contiguous-quantized attention (submit + wait).
|
|
223
|
+
* Types mirror runAttentionTieredQuant above.
|
|
224
|
+
*/
|
|
225
|
+
export declare function runAttentionContiguousQuant(
|
|
226
|
+
Q: Tensor,
|
|
227
|
+
packedK: GPUBuffer,
|
|
228
|
+
packedV: GPUBuffer,
|
|
229
|
+
scalesK: GPUBuffer,
|
|
230
|
+
scalesV: GPUBuffer,
|
|
231
|
+
numHeads: number,
|
|
232
|
+
headDim: number,
|
|
233
|
+
options?: AttentionOptions
|
|
234
|
+
): Promise<Tensor>;
|
|
235
|
+
|
|
236
|
+
/** Record contiguous-quantized attention onto a CommandRecorder. */
|
|
237
|
+
export declare function recordAttentionContiguousQuant(
|
|
238
|
+
recorder: CommandRecorder,
|
|
239
|
+
Q: Tensor,
|
|
240
|
+
packedK: GPUBuffer,
|
|
241
|
+
packedV: GPUBuffer,
|
|
242
|
+
scalesK: GPUBuffer,
|
|
243
|
+
scalesV: GPUBuffer,
|
|
244
|
+
numHeads: number,
|
|
245
|
+
headDim: number,
|
|
246
|
+
options?: AttentionOptions
|
|
247
|
+
): Promise<Tensor>;
|
|
@@ -1,7 +1,7 @@
|
|
|
1
1
|
|
|
2
2
|
|
|
3
3
|
import { getDevice, getDeviceEpoch, getDeviceLimits, getKernelCapabilities } from '../device.js';
|
|
4
|
-
import { acquireBuffer } from '../../memory/buffer-pool.js';
|
|
4
|
+
import { acquireBuffer, releaseBuffer } from '../../memory/buffer-pool.js';
|
|
5
5
|
import { createTensor } from '../tensor.js';
|
|
6
6
|
import { KernelBase } from './kernel-base.js';
|
|
7
7
|
import { TILE_SIZES } from './constants.js';
|
|
@@ -351,17 +351,31 @@ function resolveAttentionVariant(
|
|
|
351
351
|
}
|
|
352
352
|
|
|
353
353
|
|
|
354
|
-
function
|
|
354
|
+
function resolveAttentionQueryBlockSize(tier, variant = null) {
|
|
355
|
+
if (variant) {
|
|
356
|
+
const metadataBlockSize = getKernelConfig('attention', variant).variantMetadata?.queryBlockSize;
|
|
357
|
+
if (metadataBlockSize != null) {
|
|
358
|
+
if (!Number.isInteger(metadataBlockSize) || metadataBlockSize <= 0) {
|
|
359
|
+
throw new Error(`Attention kernel "${variant}" has invalid variantMetadata.queryBlockSize=${metadataBlockSize}.`);
|
|
360
|
+
}
|
|
361
|
+
return metadataBlockSize;
|
|
362
|
+
}
|
|
363
|
+
}
|
|
364
|
+
if (tier === 'tiled_large') {
|
|
365
|
+
return TILE_SIZES.ATTENTION_LARGE_BLOCK_SIZE;
|
|
366
|
+
}
|
|
367
|
+
return TILE_SIZES.ATTENTION_SMALL_BLOCK_SIZE;
|
|
368
|
+
}
|
|
369
|
+
|
|
370
|
+
function calculateAttentionWorkgroups(tier, seqLen, numHeads, variant = null) {
|
|
355
371
|
if (tier === 'subgroup') {
|
|
356
372
|
return numHeads;
|
|
357
373
|
}
|
|
358
374
|
if (tier === 'streaming') {
|
|
359
375
|
return seqLen * numHeads;
|
|
360
376
|
}
|
|
361
|
-
|
|
362
|
-
|
|
363
|
-
}
|
|
364
|
-
return Math.ceil(seqLen / TILE_SIZES.ATTENTION_SMALL_BLOCK_SIZE) * numHeads;
|
|
377
|
+
const queryBlockSize = resolveAttentionQueryBlockSize(tier, variant);
|
|
378
|
+
return Math.ceil(seqLen / queryBlockSize) * numHeads;
|
|
365
379
|
}
|
|
366
380
|
|
|
367
381
|
|
|
@@ -561,7 +575,7 @@ function resolveAttentionPlan(
|
|
|
561
575
|
caps,
|
|
562
576
|
sharedLimit
|
|
563
577
|
);
|
|
564
|
-
const workgroups = calculateAttentionWorkgroups(adaptiveSelection.tier, seqLen, numHeads);
|
|
578
|
+
const workgroups = calculateAttentionWorkgroups(adaptiveSelection.tier, seqLen, numHeads, adaptiveVariant);
|
|
565
579
|
logKernelSelectionOnce('attention', {
|
|
566
580
|
variant: adaptiveVariant,
|
|
567
581
|
reason: `path_override_fallback:${adaptiveSelection.tier}`,
|
|
@@ -603,7 +617,7 @@ function resolveAttentionPlan(
|
|
|
603
617
|
}
|
|
604
618
|
|
|
605
619
|
const tier = inferAttentionTierFromVariant(variantOverride);
|
|
606
|
-
const workgroups = calculateAttentionWorkgroups(tier, seqLen, numHeads);
|
|
620
|
+
const workgroups = calculateAttentionWorkgroups(tier, seqLen, numHeads, variantOverride);
|
|
607
621
|
logKernelSelectionOnce('attention', {
|
|
608
622
|
variant: variantOverride,
|
|
609
623
|
reason: `${selectionReason}:${tier}`,
|
|
@@ -635,7 +649,7 @@ function resolveAttentionPlan(
|
|
|
635
649
|
kvLen,
|
|
636
650
|
sharedLimit
|
|
637
651
|
);
|
|
638
|
-
const workgroups = calculateAttentionWorkgroups(tier, seqLen, numHeads);
|
|
652
|
+
const workgroups = calculateAttentionWorkgroups(tier, seqLen, numHeads, variant);
|
|
639
653
|
|
|
640
654
|
logKernelSelectionOnce('attention', {
|
|
641
655
|
variant: validatedVariant,
|
|
@@ -1039,6 +1053,79 @@ async function executeAttention(
|
|
|
1039
1053
|
);
|
|
1040
1054
|
}
|
|
1041
1055
|
|
|
1056
|
+
// ORT-style single-pass flash attention (adapted from microsoft/onnxruntime
|
|
1057
|
+
// flash_attention.wgsl.template). Single kernel, no reduce pass — each WG
|
|
1058
|
+
// handles one (head, query-tile) and processes all K online. Gated by
|
|
1059
|
+
// useOrtFlashPrefill; takes precedence over the split+reduce flash path.
|
|
1060
|
+
if (
|
|
1061
|
+
options.useOrtFlashPrefill === true
|
|
1062
|
+
&& headDim === FLASH_HEAD_DIM
|
|
1063
|
+
&& seqLen > 1
|
|
1064
|
+
&& kvLayout === 'contiguous'
|
|
1065
|
+
&& bidirectionalSpanLength === 0
|
|
1066
|
+
&& indirectBuffer == null
|
|
1067
|
+
&& K?.dtype === 'f16'
|
|
1068
|
+
&& V?.dtype === 'f16'
|
|
1069
|
+
) {
|
|
1070
|
+
return executeOrtFlashAttentionPrefill(recorder, Q, K, V, numHeads, headDim, {
|
|
1071
|
+
seqLen,
|
|
1072
|
+
kvLen,
|
|
1073
|
+
numKVHeads,
|
|
1074
|
+
scale,
|
|
1075
|
+
causal,
|
|
1076
|
+
startPos,
|
|
1077
|
+
outputBuffer,
|
|
1078
|
+
attnSoftcap,
|
|
1079
|
+
slidingWindow,
|
|
1080
|
+
kvLenBuffer,
|
|
1081
|
+
kvStart,
|
|
1082
|
+
kvLayout,
|
|
1083
|
+
kvPageTable,
|
|
1084
|
+
kvPageSize,
|
|
1085
|
+
});
|
|
1086
|
+
}
|
|
1087
|
+
|
|
1088
|
+
// Flash-attention prefill path: raises RDNA3 occupancy via KV-axis workgroup
|
|
1089
|
+
// splitting + online-softmax reduction. Gated by options.useFlashPrefill so
|
|
1090
|
+
// callers opt in deliberately (runtime config flag). Conservative conditions:
|
|
1091
|
+
// head_dim=256, prefill (seqLen>1), contiguous KV, no bidirectional span.
|
|
1092
|
+
if (!globalThis.__DOPPLER_FLASH_TRACE2__ && options.useFlashPrefill === true) {
|
|
1093
|
+
globalThis.__DOPPLER_FLASH_TRACE2__ = true;
|
|
1094
|
+
if (typeof process !== 'undefined' && process?.stderr?.write) process.stderr.write('[FLASH2] useFlash=' + options.useFlashPrefill
|
|
1095
|
+
+ ' hd=' + headDim + '/' + FLASH_HEAD_DIM
|
|
1096
|
+
+ ' seq=' + seqLen + ' kvL=' + kvLayout
|
|
1097
|
+
+ ' biDir=' + bidirectionalSpanLength
|
|
1098
|
+
+ ' indirect=' + (indirectBuffer == null ? 'n' : 'y')
|
|
1099
|
+
+ ' K=' + K?.dtype + ' V=' + V?.dtype + '\n');
|
|
1100
|
+
}
|
|
1101
|
+
if (
|
|
1102
|
+
options.useFlashPrefill === true
|
|
1103
|
+
&& headDim === FLASH_HEAD_DIM
|
|
1104
|
+
&& seqLen > 1
|
|
1105
|
+
&& kvLayout === 'contiguous'
|
|
1106
|
+
&& bidirectionalSpanLength === 0
|
|
1107
|
+
&& indirectBuffer == null
|
|
1108
|
+
&& K?.dtype === 'f16'
|
|
1109
|
+
&& V?.dtype === 'f16'
|
|
1110
|
+
) {
|
|
1111
|
+
return executeFlashAttentionPrefill(recorder, Q, K, V, numHeads, headDim, {
|
|
1112
|
+
seqLen,
|
|
1113
|
+
kvLen,
|
|
1114
|
+
numKVHeads,
|
|
1115
|
+
scale,
|
|
1116
|
+
causal,
|
|
1117
|
+
startPos,
|
|
1118
|
+
outputBuffer,
|
|
1119
|
+
attnSoftcap,
|
|
1120
|
+
slidingWindow,
|
|
1121
|
+
kvLenBuffer,
|
|
1122
|
+
kvStart,
|
|
1123
|
+
kvLayout,
|
|
1124
|
+
kvPageTable,
|
|
1125
|
+
kvPageSize,
|
|
1126
|
+
});
|
|
1127
|
+
}
|
|
1128
|
+
|
|
1042
1129
|
const limits = getDeviceLimits();
|
|
1043
1130
|
const sharedLimit = limits?.maxComputeWorkgroupStorageSize ?? Infinity;
|
|
1044
1131
|
const caps = getKernelCapabilities();
|
|
@@ -1155,6 +1242,353 @@ async function executeAttention(
|
|
|
1155
1242
|
return createTensor(outputBuf, outputDtype, [seqLen, numHeads, headDim], 'attention_output');
|
|
1156
1243
|
}
|
|
1157
1244
|
|
|
1245
|
+
// -----------------------------------------------------------------------------
|
|
1246
|
+
// Flash-attention prefill path (head_dim = 256, f16 KV)
|
|
1247
|
+
// -----------------------------------------------------------------------------
|
|
1248
|
+
// Two-pass kernel to raise RDNA3 occupancy: pass 1 processes one KV slice per
|
|
1249
|
+
// workgroup and writes per-split (acc, m, l) partials; pass 2 merges across
|
|
1250
|
+
// splits with online softmax. Single recorder, two dispatches — queue order
|
|
1251
|
+
// handles the read-after-write between passes.
|
|
1252
|
+
|
|
1253
|
+
const FLASH_BLOCK_SIZE = 32;
|
|
1254
|
+
const FLASH_HEAD_DIM = 256;
|
|
1255
|
+
const FLASH_HEAD_DIM_VECS = 64;
|
|
1256
|
+
const FLASH_REDUCE_WG = 64;
|
|
1257
|
+
|
|
1258
|
+
let flashPrefillKernel = null;
|
|
1259
|
+
let flashReduceKernel = null;
|
|
1260
|
+
|
|
1261
|
+
class FlashAttentionPrefillKernel extends KernelBase {
|
|
1262
|
+
async getPipeline(variant) {
|
|
1263
|
+
return this.getPipelineFor('attention', variant);
|
|
1264
|
+
}
|
|
1265
|
+
record(recorder, pipeline, bindGroup, workgroups) {
|
|
1266
|
+
this.recordKernel(recorder, pipeline, bindGroup, workgroups, 'attention');
|
|
1267
|
+
}
|
|
1268
|
+
dispatch(pipeline, bindGroup, workgroups) {
|
|
1269
|
+
this.dispatchKernel(pipeline, bindGroup, workgroups, 'attention');
|
|
1270
|
+
}
|
|
1271
|
+
}
|
|
1272
|
+
|
|
1273
|
+
function getFlashPrefillKernel(device) {
|
|
1274
|
+
if (!flashPrefillKernel) {
|
|
1275
|
+
flashPrefillKernel = new FlashAttentionPrefillKernel(device);
|
|
1276
|
+
}
|
|
1277
|
+
return flashPrefillKernel;
|
|
1278
|
+
}
|
|
1279
|
+
|
|
1280
|
+
function getFlashReduceKernel(device) {
|
|
1281
|
+
if (!flashReduceKernel) {
|
|
1282
|
+
flashReduceKernel = new FlashAttentionPrefillKernel(device);
|
|
1283
|
+
}
|
|
1284
|
+
return flashReduceKernel;
|
|
1285
|
+
}
|
|
1286
|
+
|
|
1287
|
+
function createFlashAttentionUniformBuffer(device, recorder, params) {
|
|
1288
|
+
// Layout mirrors the flash kernel's Uniforms struct (see
|
|
1289
|
+
// attention_prefill_flash_head256_f16kv.wgsl). 64 bytes total.
|
|
1290
|
+
return createUniformBufferWithView(
|
|
1291
|
+
'attention_flash_uniforms',
|
|
1292
|
+
64,
|
|
1293
|
+
(view) => {
|
|
1294
|
+
view.setUint32(0, params.numHeads, true);
|
|
1295
|
+
view.setUint32(4, params.numKVHeads, true);
|
|
1296
|
+
view.setUint32(8, params.headDim, true);
|
|
1297
|
+
view.setUint32(12, params.kvLen, true);
|
|
1298
|
+
view.setUint32(16, params.seqLen, true);
|
|
1299
|
+
view.setFloat32(20, params.scale, true);
|
|
1300
|
+
view.setUint32(24, params.causal ? 1 : 0, true);
|
|
1301
|
+
view.setUint32(28, params.startPos, true);
|
|
1302
|
+
view.setFloat32(32, params.attnSoftcap, true);
|
|
1303
|
+
view.setUint32(36, params.slidingWindow, true);
|
|
1304
|
+
view.setUint32(40, params.kvLenSource, true);
|
|
1305
|
+
view.setUint32(44, params.kvStart ?? 0, true);
|
|
1306
|
+
view.setUint32(48, params.pageSize ?? 0, true);
|
|
1307
|
+
view.setUint32(52, params.kvLayout ?? 0, true);
|
|
1308
|
+
view.setUint32(56, params.numKvSplits, true);
|
|
1309
|
+
view.setUint32(60, 0, true);
|
|
1310
|
+
},
|
|
1311
|
+
recorder,
|
|
1312
|
+
device
|
|
1313
|
+
);
|
|
1314
|
+
}
|
|
1315
|
+
|
|
1316
|
+
function createFlashReduceUniformBuffer(device, recorder, params) {
|
|
1317
|
+
return createUniformBufferWithView(
|
|
1318
|
+
'attention_flash_reduce_uniforms',
|
|
1319
|
+
16,
|
|
1320
|
+
(view) => {
|
|
1321
|
+
view.setUint32(0, params.numHeads, true);
|
|
1322
|
+
view.setUint32(4, params.queryLen, true);
|
|
1323
|
+
view.setUint32(8, params.numKvSplits, true);
|
|
1324
|
+
view.setUint32(12, 0, true);
|
|
1325
|
+
},
|
|
1326
|
+
recorder,
|
|
1327
|
+
device
|
|
1328
|
+
);
|
|
1329
|
+
}
|
|
1330
|
+
|
|
1331
|
+
function chooseFlashNumKvSplits(kvLen) {
|
|
1332
|
+
// Target roughly 32 workgroups × num_heads × num_kv_splits ≈ 4x RDNA3
|
|
1333
|
+
// compute-unit count. Keep at least 2× FLASH_BLOCK_SIZE KV positions per
|
|
1334
|
+
// split so each workgroup has enough work to amortise dispatch overhead.
|
|
1335
|
+
// Short prefills (kvLen ≤ 2 × BLOCK_SIZE) take the single-split fast path
|
|
1336
|
+
// which skips the reduce pass entirely.
|
|
1337
|
+
if (kvLen <= 2 * FLASH_BLOCK_SIZE) return 1;
|
|
1338
|
+
const maxSplits = Math.min(8, Math.floor(kvLen / (2 * FLASH_BLOCK_SIZE)));
|
|
1339
|
+
return Math.max(1, maxSplits);
|
|
1340
|
+
}
|
|
1341
|
+
|
|
1342
|
+
async function executeFlashAttentionPrefill(recorder, Q, K, V, numHeads, headDim, options = {}) {
|
|
1343
|
+
if (headDim !== FLASH_HEAD_DIM) {
|
|
1344
|
+
throw new Error(`[FlashAttention] headDim must be ${FLASH_HEAD_DIM}, got ${headDim}.`);
|
|
1345
|
+
}
|
|
1346
|
+
const execution = resolveAttentionExecution(recorder);
|
|
1347
|
+
const {
|
|
1348
|
+
seqLen = 1,
|
|
1349
|
+
kvLen = seqLen,
|
|
1350
|
+
numKVHeads = numHeads,
|
|
1351
|
+
scale = 1.0 / Math.sqrt(headDim),
|
|
1352
|
+
causal = true,
|
|
1353
|
+
startPos = 0,
|
|
1354
|
+
outputBuffer = null,
|
|
1355
|
+
attnSoftcap = 0,
|
|
1356
|
+
slidingWindow = 0,
|
|
1357
|
+
kvLenBuffer = null,
|
|
1358
|
+
kvStart = 0,
|
|
1359
|
+
kvLayout = 'contiguous',
|
|
1360
|
+
kvPageTable = null,
|
|
1361
|
+
kvPageSize = 0,
|
|
1362
|
+
} = options;
|
|
1363
|
+
|
|
1364
|
+
if (kvLayout !== 'contiguous') {
|
|
1365
|
+
throw new Error(`[FlashAttention] kvLayout must be "contiguous", got "${kvLayout}".`);
|
|
1366
|
+
}
|
|
1367
|
+
|
|
1368
|
+
const device = execution.device;
|
|
1369
|
+
const numQueryBlocks = Math.max(1, Math.ceil(seqLen / FLASH_BLOCK_SIZE));
|
|
1370
|
+
const numKvSplits = chooseFlashNumKvSplits(kvLen);
|
|
1371
|
+
const singleSplit = numKvSplits === 1;
|
|
1372
|
+
|
|
1373
|
+
// Final output buffer (always allocated; bound to slot 4 on the single-split
|
|
1374
|
+
// fast path where the kernel writes normalised output directly).
|
|
1375
|
+
const paddedHiddenSize = padToQ4KBlock(numHeads * headDim);
|
|
1376
|
+
const outputSize = seqLen * paddedHiddenSize * 4;
|
|
1377
|
+
const outputBuf = outputBuffer || acquireBuffer(outputSize, undefined, 'attention_flash_output');
|
|
1378
|
+
|
|
1379
|
+
// Intermediate buffers for the multi-split path. On the single-split fast
|
|
1380
|
+
// path we bypass the reduce pass entirely and bind the output buffer in
|
|
1381
|
+
// slot 4 (partial_acc slot) while m/l bindings get tiny stub buffers — the
|
|
1382
|
+
// kernel skips writes to them.
|
|
1383
|
+
const partialAccBytes = singleSplit
|
|
1384
|
+
? 4
|
|
1385
|
+
: numQueryBlocks * numHeads * numKvSplits * FLASH_BLOCK_SIZE * FLASH_HEAD_DIM * 4;
|
|
1386
|
+
const partialStatsBytes = singleSplit
|
|
1387
|
+
? 4
|
|
1388
|
+
: numQueryBlocks * numHeads * numKvSplits * FLASH_BLOCK_SIZE * 4;
|
|
1389
|
+
const partialAcc = singleSplit
|
|
1390
|
+
? outputBuf
|
|
1391
|
+
: acquireBuffer(partialAccBytes, GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST, 'flash_partial_acc');
|
|
1392
|
+
const partialM = acquireBuffer(
|
|
1393
|
+
partialStatsBytes,
|
|
1394
|
+
GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST,
|
|
1395
|
+
'flash_partial_m'
|
|
1396
|
+
);
|
|
1397
|
+
const partialL = acquireBuffer(
|
|
1398
|
+
partialStatsBytes,
|
|
1399
|
+
GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST,
|
|
1400
|
+
'flash_partial_l'
|
|
1401
|
+
);
|
|
1402
|
+
|
|
1403
|
+
// Pass 1 uniforms + dispatch.
|
|
1404
|
+
const flashUniform = createFlashAttentionUniformBuffer(device, execution.recorder, {
|
|
1405
|
+
numHeads,
|
|
1406
|
+
numKVHeads,
|
|
1407
|
+
headDim,
|
|
1408
|
+
kvLen,
|
|
1409
|
+
seqLen,
|
|
1410
|
+
scale,
|
|
1411
|
+
causal,
|
|
1412
|
+
startPos,
|
|
1413
|
+
attnSoftcap,
|
|
1414
|
+
slidingWindow,
|
|
1415
|
+
kvLenSource: kvLenBuffer ? 1 : 0,
|
|
1416
|
+
kvStart,
|
|
1417
|
+
pageSize: kvPageSize,
|
|
1418
|
+
kvLayout: 0, // contiguous only for now
|
|
1419
|
+
numKvSplits,
|
|
1420
|
+
});
|
|
1421
|
+
|
|
1422
|
+
const flashKernel = getFlashPrefillKernel(device);
|
|
1423
|
+
const flashPipeline = await flashKernel.getPipeline('prefill_flash_head256_f16kv');
|
|
1424
|
+
const kvLenBinding = kvLenBuffer || getKvLenFallbackBuffer(device);
|
|
1425
|
+
const pageTableBinding = kvPageTable || getPageTableFallbackBuffer(device);
|
|
1426
|
+
|
|
1427
|
+
const flashBindGroup = device.createBindGroup({
|
|
1428
|
+
label: 'attention_flash_prefill_bg',
|
|
1429
|
+
layout: flashPipeline.getBindGroupLayout(0),
|
|
1430
|
+
entries: [
|
|
1431
|
+
{ binding: 0, resource: { buffer: flashUniform } },
|
|
1432
|
+
{ binding: 1, resource: { buffer: Q.buffer } },
|
|
1433
|
+
{ binding: 2, resource: { buffer: K.buffer } },
|
|
1434
|
+
{ binding: 3, resource: { buffer: V.buffer } },
|
|
1435
|
+
{ binding: 4, resource: { buffer: partialAcc } },
|
|
1436
|
+
{ binding: 5, resource: { buffer: partialM } },
|
|
1437
|
+
{ binding: 6, resource: { buffer: partialL } },
|
|
1438
|
+
{ binding: 7, resource: { buffer: kvLenBinding } },
|
|
1439
|
+
{ binding: 8, resource: { buffer: pageTableBinding } },
|
|
1440
|
+
],
|
|
1441
|
+
});
|
|
1442
|
+
|
|
1443
|
+
dispatchAttentionKernel(
|
|
1444
|
+
execution,
|
|
1445
|
+
flashKernel,
|
|
1446
|
+
flashPipeline,
|
|
1447
|
+
flashBindGroup,
|
|
1448
|
+
numQueryBlocks * numHeads * numKvSplits
|
|
1449
|
+
);
|
|
1450
|
+
releaseAttentionUniform(execution, flashUniform);
|
|
1451
|
+
|
|
1452
|
+
// Pass 2 — reduce. Skipped on the single-split fast path where pass 1
|
|
1453
|
+
// already wrote the final normalised output directly to outputBuf.
|
|
1454
|
+
if (!singleSplit) {
|
|
1455
|
+
const reduceUniform = createFlashReduceUniformBuffer(device, execution.recorder, {
|
|
1456
|
+
numHeads,
|
|
1457
|
+
queryLen: seqLen,
|
|
1458
|
+
numKvSplits,
|
|
1459
|
+
});
|
|
1460
|
+
|
|
1461
|
+
const reduceKernel = getFlashReduceKernel(device);
|
|
1462
|
+
const reducePipeline = await reduceKernel.getPipeline('prefill_flash_reduce');
|
|
1463
|
+
|
|
1464
|
+
const reduceBindGroup = device.createBindGroup({
|
|
1465
|
+
label: 'attention_flash_reduce_bg',
|
|
1466
|
+
layout: reducePipeline.getBindGroupLayout(0),
|
|
1467
|
+
entries: [
|
|
1468
|
+
{ binding: 0, resource: { buffer: reduceUniform } },
|
|
1469
|
+
{ binding: 1, resource: { buffer: partialAcc } },
|
|
1470
|
+
{ binding: 2, resource: { buffer: partialM } },
|
|
1471
|
+
{ binding: 3, resource: { buffer: partialL } },
|
|
1472
|
+
{ binding: 4, resource: { buffer: outputBuf } },
|
|
1473
|
+
],
|
|
1474
|
+
});
|
|
1475
|
+
|
|
1476
|
+
const totalQh = seqLen * numHeads;
|
|
1477
|
+
const reduceWgX = Math.ceil(totalQh / FLASH_REDUCE_WG);
|
|
1478
|
+
dispatchAttentionKernel(
|
|
1479
|
+
execution,
|
|
1480
|
+
reduceKernel,
|
|
1481
|
+
reducePipeline,
|
|
1482
|
+
reduceBindGroup,
|
|
1483
|
+
[reduceWgX, FLASH_HEAD_DIM_VECS, 1]
|
|
1484
|
+
);
|
|
1485
|
+
releaseAttentionUniform(execution, reduceUniform);
|
|
1486
|
+
}
|
|
1487
|
+
|
|
1488
|
+
// Release intermediate buffers via the recorder's deferred cleanup so GPU
|
|
1489
|
+
// work completes before they re-enter the pool. On the single-split path
|
|
1490
|
+
// partialAcc IS the output buffer, so we skip it.
|
|
1491
|
+
const intermediates = singleSplit ? [partialM, partialL] : [partialAcc, partialM, partialL];
|
|
1492
|
+
if (execution.recorder) {
|
|
1493
|
+
for (const buf of intermediates) {
|
|
1494
|
+
execution.recorder.trackTemporaryBuffer(buf);
|
|
1495
|
+
}
|
|
1496
|
+
} else {
|
|
1497
|
+
device.queue.onSubmittedWorkDone().then(() => {
|
|
1498
|
+
for (const buf of intermediates) {
|
|
1499
|
+
releaseBuffer(buf);
|
|
1500
|
+
}
|
|
1501
|
+
});
|
|
1502
|
+
}
|
|
1503
|
+
|
|
1504
|
+
return createTensor(outputBuf, 'f32', [seqLen, numHeads, headDim], 'attention_flash_output');
|
|
1505
|
+
}
|
|
1506
|
+
|
|
1507
|
+
// Single-pass flash-attention dispatcher (ORT-style). 7-binding contract —
|
|
1508
|
+
// same as attention_head256_f16kv. One kernel launch, no reduce pass.
|
|
1509
|
+
// Workgroups: (num_heads, ceil(seqLen / ORT_FLASH_WG), 1) with ORT_FLASH_WG=64.
|
|
1510
|
+
const ORT_FLASH_WG = 64;
|
|
1511
|
+
|
|
1512
|
+
async function executeOrtFlashAttentionPrefill(recorder, Q, K, V, numHeads, headDim, options = {}) {
|
|
1513
|
+
if (headDim !== FLASH_HEAD_DIM) {
|
|
1514
|
+
throw new Error(`[OrtFlashAttention] headDim must be ${FLASH_HEAD_DIM}, got ${headDim}.`);
|
|
1515
|
+
}
|
|
1516
|
+
const execution = resolveAttentionExecution(recorder);
|
|
1517
|
+
const {
|
|
1518
|
+
seqLen = 1,
|
|
1519
|
+
kvLen = seqLen,
|
|
1520
|
+
numKVHeads = numHeads,
|
|
1521
|
+
scale = 1.0 / Math.sqrt(headDim),
|
|
1522
|
+
causal = true,
|
|
1523
|
+
startPos = 0,
|
|
1524
|
+
outputBuffer = null,
|
|
1525
|
+
attnSoftcap = 0,
|
|
1526
|
+
slidingWindow = 0,
|
|
1527
|
+
kvLenBuffer = null,
|
|
1528
|
+
kvStart = 0,
|
|
1529
|
+
kvLayout = 'contiguous',
|
|
1530
|
+
kvPageTable = null,
|
|
1531
|
+
kvPageSize = 0,
|
|
1532
|
+
} = options;
|
|
1533
|
+
|
|
1534
|
+
if (kvLayout !== 'contiguous') {
|
|
1535
|
+
throw new Error(`[OrtFlashAttention] kvLayout must be "contiguous", got "${kvLayout}".`);
|
|
1536
|
+
}
|
|
1537
|
+
|
|
1538
|
+
const device = execution.device;
|
|
1539
|
+
const paddedHiddenSize = padToQ4KBlock(numHeads * headDim);
|
|
1540
|
+
const outputSize = seqLen * paddedHiddenSize * 4;
|
|
1541
|
+
const outputBuf = outputBuffer || acquireBuffer(outputSize, undefined, 'attention_ort_flash_output');
|
|
1542
|
+
|
|
1543
|
+
const uniform = createAttentionUniformBuffer(device, execution.recorder, {
|
|
1544
|
+
numHeads,
|
|
1545
|
+
numKVHeads,
|
|
1546
|
+
headDim,
|
|
1547
|
+
kvLen,
|
|
1548
|
+
seqLen,
|
|
1549
|
+
scale,
|
|
1550
|
+
causal,
|
|
1551
|
+
startPos,
|
|
1552
|
+
attnSoftcap,
|
|
1553
|
+
slidingWindow,
|
|
1554
|
+
kvLenSource: kvLenBuffer ? 1 : 0,
|
|
1555
|
+
kvStart,
|
|
1556
|
+
pageSize: kvPageSize,
|
|
1557
|
+
kvLayout: 0, // contiguous only
|
|
1558
|
+
});
|
|
1559
|
+
|
|
1560
|
+
const kernel = new AttentionKernel(device);
|
|
1561
|
+
const pipeline = await kernel.getPipeline('prefill_flash_ort_head256_f16kv');
|
|
1562
|
+
const kvLenBinding = kvLenBuffer || getKvLenFallbackBuffer(device);
|
|
1563
|
+
const pageTableBinding = kvPageTable || getPageTableFallbackBuffer(device);
|
|
1564
|
+
|
|
1565
|
+
const bindGroup = device.createBindGroup({
|
|
1566
|
+
label: 'attention_ort_flash_prefill_bg',
|
|
1567
|
+
layout: pipeline.getBindGroupLayout(0),
|
|
1568
|
+
entries: [
|
|
1569
|
+
{ binding: 0, resource: { buffer: uniform } },
|
|
1570
|
+
{ binding: 1, resource: { buffer: Q.buffer } },
|
|
1571
|
+
{ binding: 2, resource: { buffer: K.buffer } },
|
|
1572
|
+
{ binding: 3, resource: { buffer: V.buffer } },
|
|
1573
|
+
{ binding: 4, resource: { buffer: outputBuf } },
|
|
1574
|
+
{ binding: 5, resource: { buffer: kvLenBinding } },
|
|
1575
|
+
{ binding: 6, resource: { buffer: pageTableBinding } },
|
|
1576
|
+
],
|
|
1577
|
+
});
|
|
1578
|
+
|
|
1579
|
+
const numSeqTiles = Math.max(1, Math.ceil(seqLen / ORT_FLASH_WG));
|
|
1580
|
+
const workgroups = [numHeads, numSeqTiles, 1];
|
|
1581
|
+
|
|
1582
|
+
if (execution.recorder) {
|
|
1583
|
+
kernel.record(execution.recorder, pipeline, bindGroup, workgroups);
|
|
1584
|
+
} else {
|
|
1585
|
+
kernel.dispatch(pipeline, bindGroup, workgroups);
|
|
1586
|
+
}
|
|
1587
|
+
|
|
1588
|
+
if (uniform) releaseUniformBuffer(uniform);
|
|
1589
|
+
return createTensor(outputBuf, 'f32', [seqLen, numHeads, headDim], 'attention_ort_flash_output');
|
|
1590
|
+
}
|
|
1591
|
+
|
|
1158
1592
|
async function executeAttentionTiered(
|
|
1159
1593
|
recorder,
|
|
1160
1594
|
Q,
|
|
@@ -1,5 +1,5 @@
|
|
|
1
1
|
// AUTO-GENERATED from src/gpu/kernels/attention_decode.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 (decode, f16 QKV + f16 output)
|
|
4
4
|
//
|
|
5
5
|
// Same as attention_decode_f16kv.wgsl but Q/K/V are f16 and output is f16.
|
|
@@ -1,5 +1,5 @@
|
|
|
1
1
|
// AUTO-GENERATED from src/gpu/kernels/attention_decode.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 (f16 KV)
|
|
4
4
|
//
|
|
5
5
|
// Same as attention.wgsl but K/V are stored as f16 for KV-cache compression.
|