doppler-gpu 0.4.7 → 0.4.8
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 +64 -0
- package/README.md +93 -144
- package/models/catalog.json +1575 -0
- package/package.json +37 -12
- package/src/cli/doppler-cli.js +28 -7
- package/src/client/doppler-registry.json +2 -2
- package/src/client/runtime/index.js +25 -2
- package/src/client/runtime/model-session.d.ts +31 -1
- package/src/client/runtime/model-session.js +31 -0
- package/src/client/runtime/model-source.d.ts +19 -2
- package/src/client/runtime/model-source.js +45 -5
- package/src/config/conversion/gemma3/translategemma-4b-1b-enes-q4k-ehf16-af32.json +21 -5
- package/src/config/conversion/gemma3/translategemma-4b-it-q4k-ehf16-af32.json +6 -1
- package/src/config/conversion/gemma4/gemma-4-e2b-it-q4k-ehf16-af16-int4ple.json +1 -1
- package/src/config/conversion/qwen3/qwen-3-reranker-0-6b-q4k-ehf16-af32.json +9 -1
- package/src/config/kernel-path-loader.js +1 -0
- package/src/config/kernels/kernel-ref-digests.js +26 -17
- package/src/config/kernels/registry.json +448 -27
- package/src/config/merge.js +11 -0
- package/src/config/param-validator.js +25 -0
- package/src/config/runtime/default.json +3 -1
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-attn128-candidate.json +20 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-maxbatch64-candidate.json +52 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-normed-ffn-candidate.json +18 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-overlap-ringtokens1-candidate.json +52 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-parity.json +85 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-sequential-candidate.json +52 -0
- package/src/config/runtime/profiles/qwen-3-5-0-8b-metal-throughput.json +85 -0
- package/src/config/runtime/profiles/qwen-3-5-2b-f16-primary-throughput.json +23 -0
- package/src/config/runtime/profiles/qwen-3-5-2b-metal-parity.json +38 -0
- package/src/config/runtime/profiles/qwen-3-5-2b-metal-throughput.json +135 -0
- package/src/config/runtime/profiles/qwen-3-embedding-0-6b-metal-stability.json +37 -0
- package/src/config/runtime/profiles/qwen-3-embedding-0-6b-metal-throughput.json +21 -0
- package/src/config/runtime/profiles/qwen-3-reranker-0-6b-metal-stability.json +41 -0
- package/src/config/runtime/profiles/qwen-3-reranker-0-6b-metal-throughput.json +41 -0
- package/src/config/runtime/profiles/rerank-stability.json +43 -0
- package/src/config/runtime/profiles/rerank-throughput.json +40 -0
- package/src/config/runtime/profiles/translategemma-1b-nativekd2-q4k-rdna3-throughput-probe.json +22 -0
- package/src/config/schema/doppler.schema.js +40 -5
- package/src/config/schema/execution-v1.schema.d.ts +41 -0
- package/src/config/schema/execution-v1.schema.js +10 -0
- package/src/config/schema/index.d.ts +3 -1
- package/src/config/schema/speculation-self.schema.d.ts +16 -0
- package/src/config/transforms/capability-transform-resolver.js +7 -1
- package/src/config/transforms/execution-graph-transforms.js +2 -2
- package/src/debug/config.d.ts +4 -0
- package/src/experimental/orchestration/index.d.ts +10 -1
- package/src/experimental/training/suite.js +1 -1
- package/src/formats/rdrr/classification.js +3 -0
- package/src/formats/rdrr/integrity.d.ts +0 -2
- package/src/formats/rdrr/integrity.js +0 -2
- package/src/formats/rdrr/types.d.ts +1 -1
- package/src/gpu/kernels/attention.d.ts +2 -0
- package/src/gpu/kernels/attention.js +75 -16
- package/src/gpu/kernels/attention_decode_online_head256_f16kv_output_gate.wgsl +268 -0
- package/src/gpu/kernels/attention_head128_f16kv.wgsl +269 -0
- package/src/gpu/kernels/fused_ffn.d.ts +32 -0
- package/src/gpu/kernels/fused_ffn.js +540 -19
- package/src/gpu/kernels/fused_ffn.wgsl +2 -1
- package/src/gpu/kernels/fused_ffn_f16.wgsl +2 -1
- package/src/gpu/kernels/fused_ffn_f16w_f32a_batched.wgsl +135 -0
- package/src/gpu/kernels/fused_ffn_q4k.wgsl +5 -3
- package/src/gpu/kernels/fused_ffn_q4k_f16.wgsl +2 -1
- package/src/gpu/kernels/fused_ffn_q4k_metal_simd16.wgsl +183 -0
- package/src/gpu/kernels/fused_gate_up_gelu_gated_f16.wgsl +2 -1
- package/src/gpu/kernels/fused_matmul_q4.wgsl +15 -6
- package/src/gpu/kernels/fused_matmul_q4_metal_simd16.wgsl +144 -0
- package/src/gpu/kernels/index.d.ts +21 -0
- package/src/gpu/kernels/index.js +13 -0
- package/src/gpu/kernels/linear-attention-core.js +413 -15
- package/src/gpu/kernels/lm-head-argmax.d.ts +10 -1
- package/src/gpu/kernels/lm-head-argmax.js +153 -57
- package/src/gpu/kernels/lm-head-select-logits.d.ts +23 -0
- package/src/gpu/kernels/lm-head-select-logits.js +186 -0
- package/src/gpu/kernels/lm_head_argmax_q4k.wgsl +289 -0
- package/src/gpu/kernels/lm_head_select_logits.wgsl +77 -0
- package/src/gpu/kernels/matmul-selection.js +5 -1
- package/src/gpu/kernels/residual_rmsnorm_pair.wgsl +127 -0
- package/src/gpu/kernels/rmsnorm-pair.d.ts +33 -0
- package/src/gpu/kernels/rmsnorm-pair.js +113 -0
- package/src/gpu/kernels/rmsnorm-stats.d.ts +29 -0
- package/src/gpu/kernels/rmsnorm-stats.js +250 -0
- package/src/gpu/kernels/silu.wgsl +7 -1
- package/src/gpu/kernels/silu_f16.wgsl +7 -1
- package/src/gpu/kernels/split-qkv-rmsnorm-rope-qk.js +15 -10
- package/src/gpu/kernels/split_qkv_rmsnorm_rope_qk.wgsl +37 -7
- package/src/gpu/kernels/split_qkv_rmsnorm_rope_qk_f16kv_cache.wgsl +37 -7
- package/src/gpu/tensor.d.ts +1 -0
- package/src/index.d.ts +1 -1
- package/src/index.js +1 -1
- package/src/inference/browser-harness-suite-helpers.js +1 -1
- package/src/inference/browser-harness-text-helpers.js +251 -20
- package/src/inference/browser-harness.js +221 -3
- package/src/inference/kv-cache/base.js +42 -1
- package/src/inference/pipelines/structured/json-grammar-mask.d.ts +18 -0
- package/src/inference/pipelines/text/attention/dispatch-params.js +8 -3
- package/src/inference/pipelines/text/attention/output-gate-fusion.js +37 -0
- package/src/inference/pipelines/text/attention/projections.js +174 -4
- package/src/inference/pipelines/text/attention/record.js +44 -21
- package/src/inference/pipelines/text/attention/run.js +18 -3
- package/src/inference/pipelines/text/config.d.ts +18 -0
- package/src/inference/pipelines/text/config.js +100 -0
- package/src/inference/pipelines/text/execution-plan.d.ts +2 -0
- package/src/inference/pipelines/text/execution-plan.js +37 -0
- package/src/inference/pipelines/text/execution-v1.d.ts +6 -0
- package/src/inference/pipelines/text/execution-v1.js +122 -8
- package/src/inference/pipelines/text/ffn/dense.js +239 -21
- package/src/inference/pipelines/text/ffn/standard.js +121 -38
- package/src/inference/pipelines/text/generator-decode-policy.js +1 -0
- package/src/inference/pipelines/text/generator-helpers.js +2 -0
- package/src/inference/pipelines/text/generator-runtime.js +1 -0
- package/src/inference/pipelines/text/generator-steps.js +2 -2
- package/src/inference/pipelines/text/generator.d.ts +18 -1
- package/src/inference/pipelines/text/generator.js +297 -19
- package/src/inference/pipelines/text/init-qkv-fusion.js +155 -62
- package/src/inference/pipelines/text/init.d.ts +22 -3
- package/src/inference/pipelines/text/init.js +2 -2
- package/src/inference/pipelines/text/layer.js +87 -8
- package/src/inference/pipelines/text/linear-attention-ab-fusion.js +208 -0
- package/src/inference/pipelines/text/linear-attention.d.ts +1 -0
- package/src/inference/pipelines/text/linear-attention.js +173 -16
- package/src/inference/pipelines/text/logits/gpu.js +2 -2
- package/src/inference/pipelines/text/logits/index.js +62 -1
- package/src/inference/pipelines/text/model-load.js +6 -0
- package/src/inference/pipelines/text/ops.js +22 -0
- package/src/inference/pipelines/text/per-layer-inputs.js +80 -11
- package/src/inference/pipelines/text/probes.js +9 -4
- package/src/inference/pipelines/text/state.d.ts +4 -4
- package/src/inference/pipelines/text/state.js +4 -0
- package/src/inference/pipelines/text/types.d.ts +27 -2
- package/src/inference/pipelines/text.d.ts +21 -1
- package/src/inference/pipelines/text.js +40 -14
- package/src/inference/tokenizers/bundled.js +45 -23
- package/src/inference/tokenizers/types.d.ts +7 -4
- package/src/loader/doppler-loader.js +25 -2
- package/src/loader/index.d.ts +3 -1
- package/src/loader/index.js +2 -0
- package/src/loader/loader-types.d.ts +3 -1
- package/src/loader/manifest-config.d.ts +18 -0
- package/src/loader/manifest-config.js +11 -3
- package/src/loader/shard-cache.d.ts +2 -0
- package/src/loader/shard-cache.js +29 -1
- package/src/models/gemma4.js +8 -0
- package/src/rules/inference/capability-transforms.rules.json +145 -20
- package/src/rules/inference/ffn.rules.json +33 -0
- package/src/rules/kernels/fused-ffn.rules.json +1 -1
- package/src/rules/kernels/lm-head-argmax.rules.json +10 -0
- package/src/rules/kernels/matmul.rules.json +1 -0
- package/src/rules/kernels/silu.rules.json +1 -0
- package/src/rules/rule-registry.js +2 -0
- package/src/storage/artifact-storage-context.js +36 -0
- package/src/tooling/node-browser-command-runner.js +42 -7
- package/src/tooling/opfs-cache.d.ts +21 -1
- package/src/tooling/opfs-cache.js +168 -32
- package/src/tooling-exports/storage.d.ts +1 -1
- package/src/tooling-exports/storage.js +1 -1
- package/src/types/gpu.d.ts +25 -0
- package/src/version.js +1 -1
- package/tests/kernels/browser/kernel-suite.js +1593 -0
- package/tests/kernels/browser/test-page.js +2629 -0
- package/tests/kernels/harness/benchmark.js +225 -0
- package/tests/kernels/harness/buffer-utils.js +88 -0
- package/tests/kernels/harness/tolerance.js +174 -0
- package/tests/kernels/reference/adam.js +15 -0
- package/tests/kernels/reference/attention.js +158 -0
- package/tests/kernels/reference/backward.js +96 -0
- package/tests/kernels/reference/bias-add.js +20 -0
- package/tests/kernels/reference/conv2d.js +58 -0
- package/tests/kernels/reference/cross-entropy.js +26 -0
- package/tests/kernels/reference/dequant.js +312 -0
- package/tests/kernels/reference/embed.js +15 -0
- package/tests/kernels/reference/gather.js +61 -0
- package/tests/kernels/reference/gelu.js +37 -0
- package/tests/kernels/reference/groupnorm.js +36 -0
- package/tests/kernels/reference/index.js +89 -0
- package/tests/kernels/reference/layernorm.js +22 -0
- package/tests/kernels/reference/matmul.js +57 -0
- package/tests/kernels/reference/modulate.js +24 -0
- package/tests/kernels/reference/moe-gather.js +83 -0
- package/tests/kernels/reference/pixel-shuffle.js +20 -0
- package/tests/kernels/reference/residual.js +55 -0
- package/tests/kernels/reference/rmsnorm.js +52 -0
- package/tests/kernels/reference/rope.js +79 -0
- package/tests/kernels/reference/sample.js +97 -0
- package/tests/kernels/reference/scatter-add.js +58 -0
- package/tests/kernels/reference/silu.js +52 -0
- package/tests/kernels/reference/softmax.js +86 -0
- package/tests/kernels/reference/split-qkv.js +58 -0
- package/tests/kernels/reference/topk.js +88 -0
- package/tests/kernels/reference/upsample2d.js +37 -0
- package/BRANDING.md +0 -14
- package/SECURITY.md +0 -19
- package/assets/doppler-runtime-map.svg +0 -82
- package/src/bootstrap.js +0 -30
- package/src/cli/doppler-cli.d.ts +0 -46
- package/src/cli/doppler-quickstart.d.ts +0 -45
- package/src/cli/doppler-serve.d.ts +0 -210
- package/src/cli/runtime-profiles.d.ts +0 -40
- package/src/client/doppler-api.browser.d.ts +0 -1
- package/src/client/doppler-api.browser.js +0 -25
- package/src/client/doppler-registry.d.ts +0 -54
- package/src/client/runtime/node-quickstart-cache.d.ts +0 -22
- package/src/client/runtime/source-runtime.d.ts +0 -26
- package/src/client/runtime/storage.d.ts +0 -5
- package/src/config/README.md +0 -61
- package/src/config/conversion/README.md +0 -62
- package/src/config/goal-completion-matrix.json +0 -552
- package/src/config/kernels/kernel-ref-digests.d.ts +0 -1
- package/src/config/kernels/moe/gpt-oss.paths.json +0 -49
- package/src/config/kernels/moe/mixtral.paths.json +0 -46
- package/src/config/kernels/registry.d.ts +0 -51
- package/src/config/merge-contract-check.d.ts +0 -16
- package/src/config/merge-contract-check.js +0 -319
- package/src/config/merge-helpers.d.ts +0 -60
- package/src/config/model-lanes/catalog-lane-resolver.d.ts +0 -66
- package/src/config/model-lanes/catalog-lane-resolver.js +0 -170
- package/src/config/param-categories.d.ts +0 -8
- package/src/config/param-validator.d.ts +0 -33
- package/src/config/platforms/loader.d.ts +0 -38
- package/src/config/quantization-contract-check.d.ts +0 -12
- package/src/config/quantization-contract-check.js +0 -91
- package/src/config/runtime/README.md +0 -28
- package/src/config/runtime-merge.d.ts +0 -5
- package/src/config/schema/claim-evidence-contract.schema.json +0 -93
- package/src/config/schema/command-surface-contract.schema.json +0 -81
- package/src/config/schema/goal-completion-matrix.schema.json +0 -223
- package/src/config/schema/model-artifact-contract.schema.json +0 -119
- package/src/config/schema/platform.schema.js +0 -1
- package/src/config/schema/policy-schema-registry.json +0 -31
- package/src/config/schema/subsystem-support-contract.schema.json +0 -48
- package/src/config/support-tiers/subsystems.json +0 -502
- package/src/config/supported-operations.d.ts +0 -14
- package/src/config/transforms/capability-transform-resolver.d.ts +0 -56
- package/src/config/transforms/execution-graph-transforms.d.ts +0 -306
- package/src/converter/conversion-plan.d.ts +0 -56
- package/src/converter/index.d.ts +0 -98
- package/src/converter/index.js +0 -59
- package/src/converter/parsers/diffusion.d.ts +0 -50
- package/src/converter/parsers/gguf.d.ts +0 -31
- package/src/converter/parsers/index.d.ts +0 -21
- package/src/converter/parsers/index.js +0 -12
- package/src/converter/parsers/transformer.d.ts +0 -31
- package/src/converter/quantization-info.d.ts +0 -38
- package/src/converter/rope-config.d.ts +0 -16
- package/src/converter/tokenizer-utils.d.ts +0 -12
- package/src/debug/reference/README.md +0 -114
- package/src/debug/reference/hf_attn_debug.py +0 -114
- package/src/debug/reference/hf_embed_check.py +0 -89
- package/src/debug/reference/hf_layer_out.py +0 -100
- package/src/debug/reference/hf_qwen35_linear_attn_debug.py +0 -268
- package/src/debug/reference/hf_rope_check.py +0 -116
- package/src/debug/reference/hf_weights.py +0 -75
- package/src/errors/doppler-error.d.ts +0 -20
- package/src/experimental/bridge/extension/background.d.ts +0 -14
- package/src/experimental/bridge/extension/background.js +0 -168
- package/src/experimental/bridge/extension/manifest.json +0 -34
- package/src/experimental/bridge/extension-client.d.ts +0 -114
- package/src/experimental/bridge/index.d.ts +0 -69
- package/src/experimental/bridge/node-bridge.js +0 -24
- package/src/experimental/bridge/protocol.d.ts +0 -96
- package/src/experimental/browser/gguf-importer.d.ts +0 -136
- package/src/experimental/browser/gguf-importer.js +0 -532
- package/src/experimental/browser/gguf-parser-browser.d.ts +0 -14
- package/src/experimental/browser/quantization.d.ts +0 -74
- package/src/experimental/browser/safetensors-parser-browser.d.ts +0 -193
- package/src/experimental/browser/shard-io-browser.d.ts +0 -57
- package/src/experimental/browser/shard-io-browser.js +0 -89
- package/src/experimental/browser/tensor-source-download.d.ts +0 -27
- package/src/experimental/diffusion/config/expansion-lanes.json +0 -29
- package/src/experimental/distribution/p2p-transport-contract.d.ts +0 -57
- package/src/experimental/distribution/p2p-transport.js +0 -75
- package/src/experimental/hotswap/intent-bundle.d.ts +0 -37
- package/src/experimental/hotswap/manifest.d.ts +0 -42
- package/src/experimental/hotswap/model-swap.js +0 -42
- package/src/experimental/hotswap/runtime.d.ts +0 -31
- package/src/experimental/training/README.md +0 -164
- package/src/experimental/training/artifacts.d.ts +0 -160
- package/src/experimental/training/attention-backward.d.ts +0 -30
- package/src/experimental/training/attention-forward.d.ts +0 -22
- package/src/experimental/training/attention-forward.js +0 -82
- package/src/experimental/training/autograd.d.ts +0 -51
- package/src/experimental/training/backward-kernels.js +0 -39
- package/src/experimental/training/checkpoint-watch.d.ts +0 -8
- package/src/experimental/training/checkpoint.d.ts +0 -36
- package/src/experimental/training/clip.d.ts +0 -9
- package/src/experimental/training/dataloader.d.ts +0 -8
- package/src/experimental/training/datasets/gepa-frontier.d.ts +0 -41
- package/src/experimental/training/datasets/gepa-frontier.js +0 -220
- package/src/experimental/training/datasets/index.d.ts +0 -24
- package/src/experimental/training/datasets/index.js +0 -13
- package/src/experimental/training/datasets/jsonl.d.ts +0 -11
- package/src/experimental/training/datasets/reploid.d.ts +0 -3
- package/src/experimental/training/datasets/reploid.js +0 -36
- package/src/experimental/training/datasets/teacher-traces.d.ts +0 -51
- package/src/experimental/training/datasets/teacher-traces.js +0 -279
- package/src/experimental/training/datasets/text-pairs.d.ts +0 -66
- package/src/experimental/training/datasets/token-batch.d.ts +0 -21
- package/src/experimental/training/datasets/token-batch.js +0 -52
- package/src/experimental/training/datasets/translation-pairs.d.ts +0 -34
- package/src/experimental/training/datasets/translation-pairs.js +0 -49
- package/src/experimental/training/distillation/artifacts.d.ts +0 -71
- package/src/experimental/training/distillation/checkpoint-watch.d.ts +0 -10
- package/src/experimental/training/distillation/dataset.d.ts +0 -59
- package/src/experimental/training/distillation/eval.d.ts +0 -34
- package/src/experimental/training/distillation/index.d.ts +0 -29
- package/src/experimental/training/distillation/runtime.d.ts +0 -20
- package/src/experimental/training/distillation/scoreboard.d.ts +0 -6
- package/src/experimental/training/distillation/stage-a.d.ts +0 -45
- package/src/experimental/training/distillation/stage-b.d.ts +0 -24
- package/src/experimental/training/distillation/student-fixture.d.ts +0 -27
- package/src/experimental/training/distillation/suite-data.d.ts +0 -45
- package/src/experimental/training/export.d.ts +0 -41
- package/src/experimental/training/index.d.ts +0 -74
- package/src/experimental/training/index.js +0 -63
- package/src/experimental/training/lora-pipeline.d.ts +0 -183
- package/src/experimental/training/lora.d.ts +0 -19
- package/src/experimental/training/loss-scaling.d.ts +0 -21
- package/src/experimental/training/loss.d.ts +0 -10
- package/src/experimental/training/objectives/base.d.ts +0 -58
- package/src/experimental/training/objectives/cross_entropy.d.ts +0 -18
- package/src/experimental/training/objectives/distill_kd.d.ts +0 -16
- package/src/experimental/training/objectives/distill_triplet.d.ts +0 -16
- package/src/experimental/training/objectives/index.d.ts +0 -12
- package/src/experimental/training/objectives/index.js +0 -6
- package/src/experimental/training/objectives/ul_stage1_joint.d.ts +0 -16
- package/src/experimental/training/objectives/ul_stage2_base.d.ts +0 -16
- package/src/experimental/training/operator-agent-eval.d.ts +0 -65
- package/src/experimental/training/operator-artifacts.d.ts +0 -62
- package/src/experimental/training/operator-command.d.ts +0 -5
- package/src/experimental/training/operator-eval.d.ts +0 -48
- package/src/experimental/training/operator-scoreboard.d.ts +0 -5
- package/src/experimental/training/optimizer.d.ts +0 -22
- package/src/experimental/training/runner.d.ts +0 -249
- package/src/experimental/training/suite.d.ts +0 -299
- package/src/experimental/training/tensor-factory.d.ts +0 -9
- package/src/experimental/training/trainer.d.ts +0 -89
- package/src/experimental/training/ul_dataset.d.ts +0 -47
- package/src/experimental/training/ul_schedule.d.ts +0 -6
- package/src/experimental/training/workload-packs/distill-glm52-doppler-agent-harness-gemma270m-f16-sft-tiny.json +0 -160
- package/src/experimental/training/workload-packs/distill-glm52-doppler-code-agent-gemma270m-f16-sft-tiny.json +0 -175
- package/src/experimental/training/workload-packs/distill-glm52-doppler-code-agent-sft-tiny.json +0 -175
- package/src/experimental/training/workload-packs/distill-glm52-doppler-js-json-gemma270m-f16-sft-tiny.json +0 -160
- package/src/experimental/training/workload-packs/distill-glm52-doppler-review-gemma270m-f16-sft-tiny.json +0 -162
- package/src/experimental/training/workload-packs/distill-glm52-doppler-wgsl-gemma270m-f16-sft-tiny.json +0 -160
- package/src/experimental/training/workload-packs/distill-translategemma-medium.json +0 -103
- package/src/experimental/training/workload-packs/distill-translategemma-tiny.json +0 -103
- package/src/experimental/training/workload-packs/lora-doppler-agent-harness-gemma270m-f16-tiny.json +0 -137
- package/src/experimental/training/workload-packs/lora-doppler-code-agent-gemma270m-f16-tiny.json +0 -152
- package/src/experimental/training/workload-packs/lora-doppler-code-agent-tiny.json +0 -152
- package/src/experimental/training/workload-packs/lora-doppler-js-json-gemma270m-f16-tiny.json +0 -137
- package/src/experimental/training/workload-packs/lora-doppler-review-gemma270m-f16-tiny.json +0 -139
- package/src/experimental/training/workload-packs/lora-doppler-wgsl-gemma270m-f16-tiny.json +0 -137
- package/src/experimental/training/workload-packs/lora-toy-medium.json +0 -91
- package/src/experimental/training/workload-packs/lora-toy-tiny.json +0 -91
- package/src/experimental/training/workload-packs/registry.json +0 -278
- package/src/experimental/training/workload-packs/ul-training-medium.json +0 -55
- package/src/experimental/training/workload-packs/ul-training-tiny.json +0 -55
- package/src/experimental/training/workloads.d.ts +0 -201
- package/src/formats/gguf/types.d.ts +0 -74
- package/src/formats/index.d.ts +0 -75
- package/src/formats/index.js +0 -19
- package/src/formats/litert/types.d.ts +0 -74
- package/src/formats/rdrr/distributed/parsing.d.ts +0 -6
- package/src/formats/rdrr/distributed/validation.d.ts +0 -23
- package/src/formats/rdrr/storage-descriptor.d.ts +0 -36
- package/src/formats/rdrr/tensor-config-validator.d.ts +0 -33
- package/src/formats/safetensors/types.d.ts +0 -67
- package/src/formats/tflite/types.d.ts +0 -61
- package/src/formats/tokenizer/index.d.ts +0 -5
- package/src/formats/tokenizer/index.js +0 -3
- package/src/formats/tokenizer/types.d.ts +0 -9
- package/src/gpu/kernel-runtime.d.ts +0 -13
- package/src/gpu/kernel-selection-cache.d.ts +0 -13
- package/src/gpu/kernel-selection-log.d.ts +0 -4
- package/src/gpu/kernel-tuner/benchmarks.d.ts +0 -144
- package/src/gpu/kernel-tuner/cache.d.ts +0 -55
- package/src/gpu/kernel-tuner/index.d.ts +0 -59
- package/src/gpu/kernel-tuner/tuner.d.ts +0 -82
- package/src/gpu/kernel-tuner/types.d.ts +0 -101
- package/src/gpu/kernel-tuner/types.js +0 -4
- package/src/gpu/kernel-tuner.d.ts +0 -33
- package/src/gpu/kernels/README.md +0 -166
- package/src/gpu/kernels/check-finiteness.d.ts +0 -15
- package/src/gpu/kernels/check-hot-vocab-stop.d.ts +0 -18
- package/src/gpu/kernels/check-stop.d.ts +0 -31
- package/src/gpu/kernels/codegen/patches/attention_decode_f16.from.attention_decode.diff +0 -213
- package/src/gpu/kernels/codegen/patches/attention_decode_f16kv.from.attention_decode.diff +0 -89
- package/src/gpu/kernels/codegen/patches/attention_f16.from.attention.diff +0 -62
- package/src/gpu/kernels/codegen/patches/attention_f16kv.from.attention.diff +0 -40
- package/src/gpu/kernels/codegen/patches/attention_small_f16.from.attention_small.diff +0 -226
- package/src/gpu/kernels/codegen/patches/attention_small_f16kv.from.attention_small.diff +0 -264
- package/src/gpu/kernels/codegen/patches/attention_streaming_f16.from.attention_streaming.diff +0 -71
- package/src/gpu/kernels/codegen/patches/attention_streaming_f16kv.from.attention_streaming.diff +0 -53
- package/src/gpu/kernels/codegen/patches/bias_add_f16.from.bias_add.diff +0 -54
- package/src/gpu/kernels/codegen/patches/dequant_mxfp4_expert_f16.from.dequant_mxfp4_expert.diff +0 -125
- package/src/gpu/kernels/codegen/patches/fused_ffn_f16.from.fused_ffn.diff +0 -428
- package/src/gpu/kernels/codegen/patches/fused_matmul_q4_batched_f16.from.fused_matmul_q4_batched.diff +0 -114
- package/src/gpu/kernels/codegen/patches/fused_matmul_rmsnorm_f16.from.fused_matmul_rmsnorm.diff +0 -477
- package/src/gpu/kernels/codegen/patches/fused_swiglu_f16.from.fused_swiglu.diff +0 -60
- package/src/gpu/kernels/codegen/patches/gelu_f16.from.gelu.diff +0 -41
- package/src/gpu/kernels/codegen/patches/layernorm_f16.from.layernorm.diff +0 -82
- package/src/gpu/kernels/codegen/patches/matmul_gemv_residual_f16.from.matmul_gemv_residual.diff +0 -87
- package/src/gpu/kernels/codegen/patches/moe_gather_f16.from.moe_gather.diff +0 -181
- package/src/gpu/kernels/codegen/patches/residual_f16.from.residual.diff +0 -65
- package/src/gpu/kernels/codegen/patches/rmsnorm_f16.from.rmsnorm.diff +0 -515
- package/src/gpu/kernels/codegen/patches/rope_f16.from.rope.diff +0 -292
- package/src/gpu/kernels/codegen/patches/sample_f16.from.sample.diff +0 -336
- package/src/gpu/kernels/codegen/patches/scatter_add_dynamic_f16.from.scatter_add_dynamic.diff +0 -65
- package/src/gpu/kernels/codegen/patches/silu_f16.from.silu.diff +0 -81
- package/src/gpu/kernels/codegen/patches/topk_f16.from.topk.diff +0 -231
- package/src/gpu/kernels/codegen/wgsl-patch-variants.js +0 -152
- package/src/gpu/kernels/codegen/wgsl-variants.js +0 -885
- package/src/gpu/kernels/constants.d.ts +0 -129
- package/src/gpu/kernels/dispatch.d.ts +0 -72
- package/src/gpu/kernels/gated-short-conv.d.ts +0 -63
- package/src/gpu/kernels/kernel-base.d.ts +0 -33
- package/src/gpu/kernels/linear-attention-core.d.ts +0 -42
- package/src/gpu/kernels/matmul-dispatch.d.ts +0 -49
- package/src/gpu/kernels/matmul-selection.d.ts +0 -87
- package/src/gpu/kernels/rep-penalty.d.ts +0 -21
- package/src/gpu/kernels/rule-matcher.d.ts +0 -30
- package/src/gpu/kernels/rule-registry.d.ts +0 -7
- package/src/gpu/kernels/types.js +0 -4
- package/src/gpu/perf-guards.d.ts +0 -13
- package/src/gpu/profiler.d.ts +0 -93
- package/src/gpu/readback-utils.d.ts +0 -16
- package/src/gpu/readback-utils.js +0 -49
- package/src/gpu/submit-probe.d.ts +0 -12
- package/src/index-browser.d.ts +0 -2
- package/src/index-browser.js +0 -2
- package/src/inference/README.md +0 -593
- package/src/inference/browser-harness-contract-helpers.d.ts +0 -5
- package/src/inference/browser-harness-diffusion-energy-suites.d.ts +0 -2
- package/src/inference/browser-harness-model-helpers.d.ts +0 -10
- package/src/inference/browser-harness-report-helpers.d.ts +0 -7
- package/src/inference/browser-harness-suite-helpers.d.ts +0 -40
- package/src/inference/browser-harness-text-helpers.d.ts +0 -152
- package/src/inference/model-lanes/execution-lane-resolver.d.ts +0 -61
- package/src/inference/model-lanes/execution-lane-resolver.js +0 -126
- package/src/inference/network-evolution.js +0 -79
- package/src/inference/pipelines/context.d.ts +0 -21
- package/src/inference/pipelines/diffusion/helpers.d.ts +0 -29
- package/src/inference/pipelines/diffusion/index.d.ts +0 -3
- package/src/inference/pipelines/diffusion/index.js +0 -3
- package/src/inference/pipelines/diffusion/scheduler.d.ts +0 -35
- package/src/inference/pipelines/diffusion/sd3-transformer.d.ts +0 -20
- package/src/inference/pipelines/diffusion/sd3-weights.d.ts +0 -21
- package/src/inference/pipelines/diffusion/text-encoder-gpu.d.ts +0 -82
- package/src/inference/pipelines/diffusion/text-encoder.d.ts +0 -29
- package/src/inference/pipelines/diffusion/types.js +0 -1
- package/src/inference/pipelines/diffusion/vae.d.ts +0 -20
- package/src/inference/pipelines/dream/energy-head-pipeline.d.ts +0 -29
- package/src/inference/pipelines/dream/energy-head-pipeline.js +0 -6
- package/src/inference/pipelines/dream/pipeline.d.ts +0 -17
- package/src/inference/pipelines/dream/pipeline.js +0 -8
- package/src/inference/pipelines/energy/index.d.ts +0 -1
- package/src/inference/pipelines/energy/index.js +0 -1
- package/src/inference/pipelines/energy/types.js +0 -1
- package/src/inference/pipelines/energy-head/index.d.ts +0 -6
- package/src/inference/pipelines/energy-head/index.js +0 -6
- package/src/inference/pipelines/factory.d.ts +0 -10
- package/src/inference/pipelines/index.d.ts +0 -22
- package/src/inference/pipelines/index.js +0 -19
- package/src/inference/pipelines/registry.d.ts +0 -15
- package/src/inference/pipelines/rng.d.ts +0 -2
- package/src/inference/pipelines/structured/tsv-grammar-mask.js +0 -166
- package/src/inference/pipelines/text/attention/attn-config.d.ts +0 -23
- package/src/inference/pipelines/text/attention/dispatch-params.d.ts +0 -173
- package/src/inference/pipelines/text/attention/index.d.ts +0 -24
- package/src/inference/pipelines/text/attention/output-projection.d.ts +0 -12
- package/src/inference/pipelines/text/attention/precision-contract.d.ts +0 -36
- package/src/inference/pipelines/text/attention/projections.d.ts +0 -208
- package/src/inference/pipelines/text/attention/quant-options.d.ts +0 -38
- package/src/inference/pipelines/text/attention/record.d.ts +0 -36
- package/src/inference/pipelines/text/attention/rmsnorm-fusion-gate.d.ts +0 -8
- package/src/inference/pipelines/text/attention/run.d.ts +0 -38
- package/src/inference/pipelines/text/attention/types.d.ts +0 -125
- package/src/inference/pipelines/text/attention.d.ts +0 -23
- package/src/inference/pipelines/text/attention.js +0 -12
- package/src/inference/pipelines/text/bdpa-steamroller.d.ts +0 -22
- package/src/inference/pipelines/text/buffer-types.js +0 -4
- package/src/inference/pipelines/text/debug-utils/config.d.ts +0 -144
- package/src/inference/pipelines/text/debug-utils/index.d.ts +0 -53
- package/src/inference/pipelines/text/debug-utils/logging.d.ts +0 -106
- package/src/inference/pipelines/text/debug-utils/tensor.d.ts +0 -119
- package/src/inference/pipelines/text/debug-utils/utils.d.ts +0 -77
- package/src/inference/pipelines/text/debug-utils.d.ts +0 -42
- package/src/inference/pipelines/text/drift-policy.d.ts +0 -57
- package/src/inference/pipelines/text/dtype-contract.d.ts +0 -53
- package/src/inference/pipelines/text/embed.d.ts +0 -69
- package/src/inference/pipelines/text/execution-runtime-builders.d.ts +0 -49
- package/src/inference/pipelines/text/failure-reduction.d.ts +0 -84
- package/src/inference/pipelines/text/failure-reduction.js +0 -180
- package/src/inference/pipelines/text/ffn/dense.d.ts +0 -99
- package/src/inference/pipelines/text/ffn/index.d.ts +0 -23
- package/src/inference/pipelines/text/ffn/moe.d.ts +0 -24
- package/src/inference/pipelines/text/ffn/sandwich.d.ts +0 -27
- package/src/inference/pipelines/text/ffn/standard.d.ts +0 -26
- package/src/inference/pipelines/text/ffn/types.d.ts +0 -30
- package/src/inference/pipelines/text/ffn.d.ts +0 -31
- package/src/inference/pipelines/text/ffn.js +0 -18
- package/src/inference/pipelines/text/finiteness-guard-status.d.ts +0 -11
- package/src/inference/pipelines/text/finiteness-policy.d.ts +0 -39
- package/src/inference/pipelines/text/generator-decode-policy.d.ts +0 -81
- package/src/inference/pipelines/text/generator-helpers.d.ts +0 -45
- package/src/inference/pipelines/text/generator-prefill-helpers.d.ts +0 -84
- package/src/inference/pipelines/text/generator-runtime.d.ts +0 -93
- package/src/inference/pipelines/text/generator-steps.d.ts +0 -186
- package/src/inference/pipelines/text/index.d.ts +0 -5
- package/src/inference/pipelines/text/index.js +0 -6
- package/src/inference/pipelines/text/kernel-trace.d.ts +0 -119
- package/src/inference/pipelines/text/layer.d.ts +0 -118
- package/src/inference/pipelines/text/load-errors.d.ts +0 -21
- package/src/inference/pipelines/text/logits/precision-policy.d.ts +0 -11
- package/src/inference/pipelines/text/logits/types.js +0 -4
- package/src/inference/pipelines/text/logits.d.ts +0 -27
- package/src/inference/pipelines/text/logits.js +0 -16
- package/src/inference/pipelines/text/lora-apply.d.ts +0 -28
- package/src/inference/pipelines/text/moe-cache.d.ts +0 -32
- package/src/inference/pipelines/text/moe-cpu-gptoss.d.ts +0 -9
- package/src/inference/pipelines/text/moe-cpu.d.ts +0 -13
- package/src/inference/pipelines/text/moe-gpu.d.ts +0 -26
- package/src/inference/pipelines/text/moe-helpers.d.ts +0 -12
- package/src/inference/pipelines/text/moe-shape-validator.d.ts +0 -70
- package/src/inference/pipelines/text/operator-events.d.ts +0 -125
- package/src/inference/pipelines/text/operator-identity.d.ts +0 -60
- package/src/inference/pipelines/text/ops.d.ts +0 -222
- package/src/inference/pipelines/text/per-layer-inputs.d.ts +0 -107
- package/src/inference/pipelines/text/probes.d.ts +0 -28
- package/src/inference/pipelines/text/sampling-config.d.ts +0 -48
- package/src/inference/pipelines/text/stage-names.d.ts +0 -88
- package/src/inference/pipelines/text/tsir-fixture-writer.d.ts +0 -57
- package/src/inference/pipelines/text/types.js +0 -4
- package/src/inference/tokenizers/base.d.ts +0 -44
- package/src/inference/tokenizers/behavior-flags.d.ts +0 -9
- package/src/inference/tokenizers/bpe.d.ts +0 -27
- package/src/inference/tokenizers/bundled.d.ts +0 -67
- package/src/inference/tokenizers/sentencepiece.d.ts +0 -30
- package/src/inference/tokenizers/types.js +0 -7
- package/src/loader/embedding-limit-preflight.d.ts +0 -31
- package/src/loader/per-layer-input-loader.d.ts +0 -30
- package/src/loader/tensors/functional-descriptor-loader.d.ts +0 -9
- package/src/loader/tensors/source-transform.d.ts +0 -65
- package/src/memory/address-table.d.ts +0 -96
- package/src/rules/execution-rules-contract-check.d.ts +0 -21
- package/src/rules/layer-pattern-contract-check.d.ts +0 -17
- package/src/rules/rule-registry.d.ts +0 -77
- package/src/storage/artifact-storage-context.d.ts +0 -63
- package/src/storage/backends/idb-store.d.ts +0 -52
- package/src/storage/backends/memory-store.d.ts +0 -36
- package/src/storage/backends/opfs-store.d.ts +0 -41
- package/src/storage/blake3.d.ts +0 -17
- package/src/storage/download-types.d.ts +0 -157
- package/src/storage/downloader.d.ts +0 -103
- package/src/storage/index.d.ts +0 -274
- package/src/storage/index.js +0 -201
- package/src/storage/preflight.d.ts +0 -144
- package/src/storage/preflight.js +0 -360
- package/src/storage/quickstart-downloader.d.ts +0 -157
- package/src/storage/quickstart-downloader.js +0 -289
- package/src/storage/source-artifact-store.d.ts +0 -59
- package/src/sw.js +0 -184
- package/src/tooling/command-api-constants.d.ts +0 -9
- package/src/tooling/command-api-family-normalizers.d.ts +0 -10
- package/src/tooling/command-api-helpers.d.ts +0 -23
- package/src/tooling/command-runner-shared.d.ts +0 -89
- package/src/tooling/conversion-config-materializer.d.ts +0 -14
- package/src/tooling/conversion-config-materializer.js +0 -96
- package/src/tooling/diagnose-runner.d.ts +0 -28
- package/src/tooling/hf-registry-utils.d.ts +0 -55
- package/src/tooling/kernel-path-builder/custom-runtime-facts.d.ts +0 -18
- package/src/tooling/kernel-path-builder/custom-runtime-facts.js +0 -72
- package/src/tooling/kernel-path-builder/index.d.ts +0 -65
- package/src/tooling/kernel-path-builder/index.js +0 -1361
- package/src/tooling/kernel-path-builder/runtime-overlay.d.ts +0 -9
- package/src/tooling/kernel-path-builder/runtime-overlay.js +0 -147
- package/src/tooling/lean-execution-contract-runner.d.ts +0 -33
- package/src/tooling/lean-execution-contract-runner.js +0 -158
- package/src/tooling/lean-execution-contract.d.ts +0 -16
- package/src/tooling/lean-execution-contract.js +0 -81
- package/src/tooling/litert-package-runtime.d.ts +0 -46
- package/src/tooling/node-convert-worker-pool.d.ts +0 -16
- package/src/tooling/node-convert-worker.d.ts +0 -1
- package/src/tooling/node-converter.d.ts +0 -1
- package/src/tooling/node-file-fetch.d.ts +0 -1
- package/src/tooling/node-source-runtime.d.ts +0 -24
- package/src/tooling/node-webgpu.d.ts +0 -17
- package/src/tooling/precision-replay-math.js +0 -174
- package/src/tooling/program-bundle-host.d.ts +0 -16
- package/src/tooling/program-bundle-host.js +0 -20
- package/src/tooling/runtime-input-composition.d.ts +0 -37
- package/src/tooling/source-package-profiles.d.ts +0 -50
- package/src/tooling/source-runtime-converter-config.d.ts +0 -19
- package/src/tooling/source-runtime-materializer.d.ts +0 -6
- package/src/tooling/source-runtime-materializer.js +0 -89
- package/src/tooling-exports.browser.d.ts +0 -7
- package/src/types/chrome.d.ts +0 -36
- package/src/types/chrome.js +0 -1
- package/src/types/index.d.ts +0 -1
- package/src/types/index.js +0 -1
- package/src/utils/array-buffer.d.ts +0 -4
- package/src/utils/canonical-hash.d.ts +0 -7
- package/src/utils/clone-json.d.ts +0 -1
- package/src/utils/hf-resolve-url.d.ts +0 -16
- package/src/utils/index.d.ts +0 -10
- package/src/utils/index.js +0 -10
- package/src/utils/load-json.d.ts +0 -5
- package/src/utils/plain-object.d.ts +0 -1
- package/src/utils/sha256.d.ts +0 -5
- package/src/utils/stable-sort-object.d.ts +0 -1
|
@@ -19,6 +19,7 @@ import { getKernelPathAttentionVariant, getKernelPathStrict } from '../../config
|
|
|
19
19
|
import { selectRuleValue as selectKernelRuleValue } from './rule-registry.js';
|
|
20
20
|
import { selectRuleValue as selectSharedRuleValue } from '../../rules/rule-registry.js';
|
|
21
21
|
import { logKernelSelectionOnce } from '../kernel-selection-log.js';
|
|
22
|
+
import { getRuntimeConfig } from '../../config/runtime.js';
|
|
22
23
|
|
|
23
24
|
// Track if we've logged the attention tier selection (avoid spam)
|
|
24
25
|
let loggedAttentionTier = false;
|
|
@@ -60,6 +61,7 @@ function getContiguousQuantMaxKVLen() {
|
|
|
60
61
|
let kvLenFallbackBuffer = null;
|
|
61
62
|
let kvLenFallbackBufferEpoch = -1;
|
|
62
63
|
const U32_BYTES = Uint32Array.BYTES_PER_ELEMENT;
|
|
64
|
+
const F32_BYTES = Float32Array.BYTES_PER_ELEMENT;
|
|
63
65
|
|
|
64
66
|
|
|
65
67
|
function getKvLenFallbackBuffer(device) {
|
|
@@ -93,13 +95,10 @@ function getPageTableFallbackBuffer(device) {
|
|
|
93
95
|
return pageTableFallbackBuffer;
|
|
94
96
|
}
|
|
95
97
|
|
|
96
|
-
|
|
97
|
-
|
|
98
|
-
|
|
99
98
|
class AttentionKernel extends KernelBase {
|
|
100
99
|
|
|
101
|
-
async getPipeline(variant) {
|
|
102
|
-
return this.getPipelineFor('attention', variant);
|
|
100
|
+
async getPipeline(variant, constants = null) {
|
|
101
|
+
return this.getPipelineFor('attention', variant, null, constants);
|
|
103
102
|
}
|
|
104
103
|
|
|
105
104
|
|
|
@@ -122,6 +121,36 @@ class AttentionKernel extends KernelBase {
|
|
|
122
121
|
}
|
|
123
122
|
}
|
|
124
123
|
|
|
124
|
+
function resolveOnlineDecodePipelineConstants(variant, headDim, options = {}) {
|
|
125
|
+
if (!variant?.startsWith('decode_online')) {
|
|
126
|
+
return null;
|
|
127
|
+
}
|
|
128
|
+
const policy = getRuntimeConfig()?.inference?.session?.attentionDecodeOnline;
|
|
129
|
+
const workgroupSize = policy?.workgroupSize;
|
|
130
|
+
const constants = {};
|
|
131
|
+
if (workgroupSize != null) {
|
|
132
|
+
if (workgroupSize !== 128 && workgroupSize !== 256) {
|
|
133
|
+
throw new Error(`attentionDecodeOnline.workgroupSize must be 128 or 256, got ${workgroupSize}.`);
|
|
134
|
+
}
|
|
135
|
+
if (headDim > workgroupSize * 2) {
|
|
136
|
+
throw new Error(
|
|
137
|
+
`attentionDecodeOnline.workgroupSize=${workgroupSize} cannot cover headDim=${headDim}; ` +
|
|
138
|
+
`requires headDim <= ${workgroupSize * 2}.`
|
|
139
|
+
);
|
|
140
|
+
}
|
|
141
|
+
constants.WORKGROUP_SIZE = workgroupSize;
|
|
142
|
+
}
|
|
143
|
+
if (
|
|
144
|
+
variant === 'decode_online_head256_f16kv_output_gate'
|
|
145
|
+
&& policy?.useDirectContiguousKVLayout === true
|
|
146
|
+
&& options.kvLayout === 'contiguous'
|
|
147
|
+
&& options.slidingWindow === 0
|
|
148
|
+
) {
|
|
149
|
+
constants.USE_DIRECT_KV_LAYOUT = true;
|
|
150
|
+
}
|
|
151
|
+
return Object.keys(constants).length > 0 ? constants : null;
|
|
152
|
+
}
|
|
153
|
+
|
|
125
154
|
class AttentionTieredKernel extends KernelBase {
|
|
126
155
|
|
|
127
156
|
async getPipeline(variant) {
|
|
@@ -1047,6 +1076,7 @@ async function executeAttention(
|
|
|
1047
1076
|
kvPageTable = null,
|
|
1048
1077
|
kvPageSize = 0,
|
|
1049
1078
|
kernelPath = null,
|
|
1079
|
+
outputGate = null,
|
|
1050
1080
|
} = options;
|
|
1051
1081
|
if (!Number.isFinite(bidirectionalSpanStart) || Math.floor(bidirectionalSpanStart) !== bidirectionalSpanStart || bidirectionalSpanStart < 0) {
|
|
1052
1082
|
throw new Error(`Attention bidirectionalSpanStart must be a non-negative integer, got ${bidirectionalSpanStart}.`);
|
|
@@ -1160,7 +1190,12 @@ async function executeAttention(
|
|
|
1160
1190
|
}
|
|
1161
1191
|
|
|
1162
1192
|
const kernel = new AttentionKernel(execution.device);
|
|
1163
|
-
const
|
|
1193
|
+
const pipelineConstants = resolveOnlineDecodePipelineConstants(plan.variant, headDim, {
|
|
1194
|
+
kvLayout,
|
|
1195
|
+
slidingWindow,
|
|
1196
|
+
});
|
|
1197
|
+
const usesOutputGateFusion = plan.variant === 'decode_online_head256_f16kv_output_gate';
|
|
1198
|
+
const pipeline = await kernel.getPipeline(plan.variant, pipelineConstants);
|
|
1164
1199
|
|
|
1165
1200
|
const outputConfig = getKernelConfig('attention', plan.variant);
|
|
1166
1201
|
const outputDtype = outputConfig.outputDtype;
|
|
@@ -1196,6 +1231,19 @@ async function executeAttention(
|
|
|
1196
1231
|
|
|
1197
1232
|
const kvLenBinding = kvLenBuffer || getKvLenFallbackBuffer(execution.device);
|
|
1198
1233
|
const pageTableBinding = kvPageTable || getPageTableFallbackBuffer(execution.device);
|
|
1234
|
+
const outputGateBinding = usesOutputGateFusion ? outputGate?.buffer : null;
|
|
1235
|
+
if (usesOutputGateFusion) {
|
|
1236
|
+
if (outputGate?.dtype !== 'f32') {
|
|
1237
|
+
throw new Error(`[Attention] outputGate fusion requires f32 gate tensor; got ${String(outputGate?.dtype)}.`);
|
|
1238
|
+
}
|
|
1239
|
+
const requiredGateBytes = numHeads * headDim * F32_BYTES;
|
|
1240
|
+
const actualGateBytes = outputGate?.buffer?.size;
|
|
1241
|
+
if (Number.isFinite(actualGateBytes) && actualGateBytes < requiredGateBytes) {
|
|
1242
|
+
throw new Error(
|
|
1243
|
+
`[Attention] outputGate fusion requires at least ${requiredGateBytes} gate bytes; got ${actualGateBytes}.`
|
|
1244
|
+
);
|
|
1245
|
+
}
|
|
1246
|
+
}
|
|
1199
1247
|
assertAttentionBindGroupBuffer('attention', plan.variant, 0, 'uniforms', uniformBuffer);
|
|
1200
1248
|
assertAttentionBindGroupBuffer('attention', plan.variant, 1, 'Q', Q?.buffer, [
|
|
1201
1249
|
`QLabel=${Q?.label ?? 'unknown'}`,
|
|
@@ -1214,18 +1262,27 @@ async function executeAttention(
|
|
|
1214
1262
|
assertAttentionBindGroupBuffer('attention', plan.variant, 6, 'pageTable', pageTableBinding, [
|
|
1215
1263
|
`kvLayout=${kvLayout}`,
|
|
1216
1264
|
]);
|
|
1265
|
+
if (usesOutputGateFusion) {
|
|
1266
|
+
assertAttentionBindGroupBuffer('attention', plan.variant, 7, 'outputGate', outputGateBinding, [
|
|
1267
|
+
`useOutputGateFusion=${usesOutputGateFusion}`,
|
|
1268
|
+
]);
|
|
1269
|
+
}
|
|
1270
|
+
const bindGroupEntries = [
|
|
1271
|
+
{ binding: 0, resource: { buffer: uniformBuffer } },
|
|
1272
|
+
{ binding: 1, resource: { buffer: Q.buffer } },
|
|
1273
|
+
{ binding: 2, resource: { buffer: K.buffer } },
|
|
1274
|
+
{ binding: 3, resource: { buffer: V.buffer } },
|
|
1275
|
+
{ binding: 4, resource: { buffer: outputBuf } },
|
|
1276
|
+
{ binding: 5, resource: { buffer: kvLenBinding } },
|
|
1277
|
+
{ binding: 6, resource: { buffer: pageTableBinding } },
|
|
1278
|
+
];
|
|
1279
|
+
if (usesOutputGateFusion) {
|
|
1280
|
+
bindGroupEntries.push({ binding: 7, resource: { buffer: outputGateBinding } });
|
|
1281
|
+
}
|
|
1217
1282
|
const bindGroup = execution.device.createBindGroup({
|
|
1218
1283
|
label: 'attention_bind_group',
|
|
1219
1284
|
layout: getPipelineBindGroupLayout(pipeline, 0),
|
|
1220
|
-
entries:
|
|
1221
|
-
{ binding: 0, resource: { buffer: uniformBuffer } },
|
|
1222
|
-
{ binding: 1, resource: { buffer: Q.buffer } },
|
|
1223
|
-
{ binding: 2, resource: { buffer: K.buffer } },
|
|
1224
|
-
{ binding: 3, resource: { buffer: V.buffer } },
|
|
1225
|
-
{ binding: 4, resource: { buffer: outputBuf } },
|
|
1226
|
-
{ binding: 5, resource: { buffer: kvLenBinding } },
|
|
1227
|
-
{ binding: 6, resource: { buffer: pageTableBinding } },
|
|
1228
|
-
],
|
|
1285
|
+
entries: bindGroupEntries,
|
|
1229
1286
|
});
|
|
1230
1287
|
|
|
1231
1288
|
if (!indirectBuffer && limits && plan.workgroups > limits.maxComputeWorkgroupsPerDimension) {
|
|
@@ -1247,7 +1304,9 @@ async function executeAttention(
|
|
|
1247
1304
|
|
|
1248
1305
|
releaseAttentionUniform(execution, uniformBuffer);
|
|
1249
1306
|
|
|
1250
|
-
|
|
1307
|
+
const outputTensor = createTensor(outputBuf, outputDtype, [seqLen, numHeads, headDim], 'attention_output');
|
|
1308
|
+
outputTensor.outputGateFused = usesOutputGateFusion;
|
|
1309
|
+
return outputTensor;
|
|
1251
1310
|
}
|
|
1252
1311
|
|
|
1253
1312
|
// -----------------------------------------------------------------------------
|
|
@@ -0,0 +1,268 @@
|
|
|
1
|
+
// Online Decode Attention Kernel with output-gate fusion
|
|
2
|
+
// (fixed head_dim 256, f32 Q + f16 KV + f32 output)
|
|
3
|
+
//
|
|
4
|
+
// Extends attention_decode_online_head256_f16kv.wgsl with a f32 output-gate
|
|
5
|
+
// buffer. Used only by execution graphs or runtime profiles that pin this
|
|
6
|
+
// explicit fused kernel identity.
|
|
7
|
+
|
|
8
|
+
enable f16;
|
|
9
|
+
enable subgroups;
|
|
10
|
+
|
|
11
|
+
const WORKGROUP_SIZE: u32 = 256u;
|
|
12
|
+
const MAX_SUBGROUPS: u32 = 256u;
|
|
13
|
+
const HEAD_DIM: u32 = 256u;
|
|
14
|
+
const HEAD_DIM_VECS: u32 = 64u;
|
|
15
|
+
override USE_DIRECT_KV_LAYOUT: bool = false;
|
|
16
|
+
|
|
17
|
+
struct Uniforms {
|
|
18
|
+
num_heads: u32,
|
|
19
|
+
num_kv_heads: u32,
|
|
20
|
+
head_dim: u32,
|
|
21
|
+
kv_len: u32,
|
|
22
|
+
seq_len: u32,
|
|
23
|
+
scale: f32,
|
|
24
|
+
is_causal: u32,
|
|
25
|
+
start_pos: u32,
|
|
26
|
+
attn_softcap: f32,
|
|
27
|
+
sliding_window: u32,
|
|
28
|
+
kv_len_source: u32,
|
|
29
|
+
kv_start: u32,
|
|
30
|
+
page_size: u32,
|
|
31
|
+
kv_layout: u32,
|
|
32
|
+
_pad: u32,
|
|
33
|
+
}
|
|
34
|
+
|
|
35
|
+
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
36
|
+
@group(0) @binding(1) var<storage, read> Q: array<f32>;
|
|
37
|
+
@group(0) @binding(2) var<storage, read> K: array<f16>;
|
|
38
|
+
@group(0) @binding(3) var<storage, read> V: array<f16>;
|
|
39
|
+
@group(0) @binding(4) var<storage, read_write> output: array<f32>;
|
|
40
|
+
@group(0) @binding(5) var<storage, read> kv_len_buffer: array<u32>;
|
|
41
|
+
@group(0) @binding(6) var<storage, read> page_table: array<u32>;
|
|
42
|
+
@group(0) @binding(7) var<storage, read> output_gate: array<f32>;
|
|
43
|
+
|
|
44
|
+
var<workgroup> shared_q: array<vec4<f32>, HEAD_DIM_VECS>;
|
|
45
|
+
var<workgroup> shared_scores: array<f32, WORKGROUP_SIZE>;
|
|
46
|
+
var<workgroup> shared_v_bases: array<u32, WORKGROUP_SIZE>;
|
|
47
|
+
var<workgroup> sg_max: array<f32, MAX_SUBGROUPS>;
|
|
48
|
+
var<workgroup> sg_sum: array<f32, MAX_SUBGROUPS>;
|
|
49
|
+
var<workgroup> global_max: f32;
|
|
50
|
+
var<workgroup> global_sum: f32;
|
|
51
|
+
|
|
52
|
+
fn get_kv_head_idx(query_head_idx: u32) -> u32 {
|
|
53
|
+
let heads_per_kv = u.num_heads / u.num_kv_heads;
|
|
54
|
+
return query_head_idx / heads_per_kv;
|
|
55
|
+
}
|
|
56
|
+
|
|
57
|
+
fn get_kv_len() -> u32 {
|
|
58
|
+
if (u.kv_len_source == 0u) {
|
|
59
|
+
return u.kv_len;
|
|
60
|
+
}
|
|
61
|
+
return kv_len_buffer[0];
|
|
62
|
+
}
|
|
63
|
+
|
|
64
|
+
fn is_masked(abs_key: u32) -> bool {
|
|
65
|
+
let abs_query = u.start_pos;
|
|
66
|
+
if (u.is_causal != 0u && abs_key > abs_query) { return true; }
|
|
67
|
+
if (u.sliding_window > 0u && abs_query >= u.sliding_window) {
|
|
68
|
+
if (abs_key < abs_query - u.sliding_window + 1u) { return true; }
|
|
69
|
+
}
|
|
70
|
+
return false;
|
|
71
|
+
}
|
|
72
|
+
|
|
73
|
+
fn get_kv_pos(key_pos: u32) -> u32 {
|
|
74
|
+
let abs_key = u.kv_start + key_pos;
|
|
75
|
+
if (USE_DIRECT_KV_LAYOUT) {
|
|
76
|
+
return abs_key;
|
|
77
|
+
}
|
|
78
|
+
if (u.kv_layout == 1u && u.sliding_window > 0u) {
|
|
79
|
+
return abs_key % u.sliding_window;
|
|
80
|
+
}
|
|
81
|
+
if (u.kv_layout == 2u) {
|
|
82
|
+
let page_idx = abs_key / u.page_size;
|
|
83
|
+
let in_page = abs_key - (page_idx * u.page_size);
|
|
84
|
+
let phys_page = page_table[page_idx];
|
|
85
|
+
return phys_page * u.page_size + in_page;
|
|
86
|
+
}
|
|
87
|
+
return abs_key;
|
|
88
|
+
}
|
|
89
|
+
|
|
90
|
+
fn sigmoid(x: f32) -> f32 {
|
|
91
|
+
return 1.0 / (1.0 + exp(-x));
|
|
92
|
+
}
|
|
93
|
+
|
|
94
|
+
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
95
|
+
fn main(
|
|
96
|
+
@builtin(local_invocation_id) local_id: vec3<u32>,
|
|
97
|
+
@builtin(workgroup_id) workgroup_id: vec3<u32>,
|
|
98
|
+
@builtin(subgroup_size) subgroup_size: u32,
|
|
99
|
+
@builtin(subgroup_invocation_id) sg_tid: u32,
|
|
100
|
+
) {
|
|
101
|
+
if (u.head_dim != HEAD_DIM) {
|
|
102
|
+
return;
|
|
103
|
+
}
|
|
104
|
+
|
|
105
|
+
let head_idx = workgroup_id.x;
|
|
106
|
+
let tid = local_id.x;
|
|
107
|
+
let kv_len = get_kv_len();
|
|
108
|
+
let q_offset = head_idx * HEAD_DIM;
|
|
109
|
+
|
|
110
|
+
if (kv_len == 0u) {
|
|
111
|
+
output[q_offset + tid] = 0.0;
|
|
112
|
+
return;
|
|
113
|
+
}
|
|
114
|
+
|
|
115
|
+
let subgroup_id = tid / subgroup_size;
|
|
116
|
+
let num_subgroups = (WORKGROUP_SIZE + subgroup_size - 1u) / subgroup_size;
|
|
117
|
+
if (num_subgroups > MAX_SUBGROUPS) {
|
|
118
|
+
return;
|
|
119
|
+
}
|
|
120
|
+
|
|
121
|
+
let kv_head_idx = get_kv_head_idx(head_idx);
|
|
122
|
+
|
|
123
|
+
if (tid < HEAD_DIM_VECS) {
|
|
124
|
+
let d = tid * 4u;
|
|
125
|
+
shared_q[tid] = vec4<f32>(
|
|
126
|
+
Q[q_offset + d],
|
|
127
|
+
Q[q_offset + d + 1u],
|
|
128
|
+
Q[q_offset + d + 2u],
|
|
129
|
+
Q[q_offset + d + 3u]
|
|
130
|
+
);
|
|
131
|
+
}
|
|
132
|
+
workgroupBarrier();
|
|
133
|
+
|
|
134
|
+
var running_max: f32 = -3.402823e+38;
|
|
135
|
+
var running_sum: f32 = 0.0;
|
|
136
|
+
var out_accum: f32 = 0.0;
|
|
137
|
+
|
|
138
|
+
var start_k: u32 = 0u;
|
|
139
|
+
if (u.sliding_window > 0u && kv_len > u.sliding_window) {
|
|
140
|
+
start_k = kv_len - u.sliding_window;
|
|
141
|
+
start_k = (start_k / WORKGROUP_SIZE) * WORKGROUP_SIZE;
|
|
142
|
+
}
|
|
143
|
+
|
|
144
|
+
for (var k_start: u32 = start_k; k_start < kv_len; k_start = k_start + WORKGROUP_SIZE) {
|
|
145
|
+
let k_pos = k_start + tid;
|
|
146
|
+
let valid_k = k_pos < kv_len;
|
|
147
|
+
var masked = false;
|
|
148
|
+
var score: f32 = -3.402823e+38;
|
|
149
|
+
|
|
150
|
+
if (valid_k) {
|
|
151
|
+
let abs_key = u.kv_start + k_pos;
|
|
152
|
+
masked = is_masked(abs_key);
|
|
153
|
+
let k_idx = get_kv_pos(k_pos);
|
|
154
|
+
let k_offset = k_idx * u.num_kv_heads * HEAD_DIM + kv_head_idx * HEAD_DIM;
|
|
155
|
+
if (!USE_DIRECT_KV_LAYOUT) {
|
|
156
|
+
shared_v_bases[tid] = k_offset;
|
|
157
|
+
}
|
|
158
|
+
if (!masked) {
|
|
159
|
+
var dot_accum: f32 = 0.0;
|
|
160
|
+
for (var d4: u32 = 0u; d4 < HEAD_DIM_VECS; d4 = d4 + 2u) {
|
|
161
|
+
let d0 = d4 * 4u;
|
|
162
|
+
let d1 = d0 + 4u;
|
|
163
|
+
let k0 = vec4<f32>(
|
|
164
|
+
f32(K[k_offset + d0]),
|
|
165
|
+
f32(K[k_offset + d0 + 1u]),
|
|
166
|
+
f32(K[k_offset + d0 + 2u]),
|
|
167
|
+
f32(K[k_offset + d0 + 3u])
|
|
168
|
+
);
|
|
169
|
+
let k1 = vec4<f32>(
|
|
170
|
+
f32(K[k_offset + d1]),
|
|
171
|
+
f32(K[k_offset + d1 + 1u]),
|
|
172
|
+
f32(K[k_offset + d1 + 2u]),
|
|
173
|
+
f32(K[k_offset + d1 + 3u])
|
|
174
|
+
);
|
|
175
|
+
dot_accum = dot_accum + dot(shared_q[d4], k0) + dot(shared_q[d4 + 1u], k1);
|
|
176
|
+
}
|
|
177
|
+
score = dot_accum * u.scale;
|
|
178
|
+
if (u.attn_softcap > 0.0) {
|
|
179
|
+
score = tanh(score / u.attn_softcap) * u.attn_softcap;
|
|
180
|
+
}
|
|
181
|
+
}
|
|
182
|
+
}
|
|
183
|
+
|
|
184
|
+
let chunk_max = subgroupMax(score);
|
|
185
|
+
if (sg_tid == 0u && subgroup_id < num_subgroups) {
|
|
186
|
+
sg_max[subgroup_id] = chunk_max;
|
|
187
|
+
}
|
|
188
|
+
workgroupBarrier();
|
|
189
|
+
|
|
190
|
+
if (tid == 0u) {
|
|
191
|
+
var m: f32 = -3.402823e+38;
|
|
192
|
+
for (var s: u32 = 0u; s < num_subgroups; s++) {
|
|
193
|
+
m = max(m, sg_max[s]);
|
|
194
|
+
}
|
|
195
|
+
global_max = m;
|
|
196
|
+
}
|
|
197
|
+
workgroupBarrier();
|
|
198
|
+
|
|
199
|
+
let chunk_max_val = global_max;
|
|
200
|
+
let new_max = max(running_max, chunk_max_val);
|
|
201
|
+
let rescale = exp(running_max - new_max);
|
|
202
|
+
|
|
203
|
+
var exp_score: f32 = 0.0;
|
|
204
|
+
if (valid_k && !masked) {
|
|
205
|
+
exp_score = exp(score - new_max);
|
|
206
|
+
}
|
|
207
|
+
shared_scores[tid] = exp_score;
|
|
208
|
+
|
|
209
|
+
let chunk_sum = subgroupAdd(exp_score);
|
|
210
|
+
if (sg_tid == 0u && subgroup_id < num_subgroups) {
|
|
211
|
+
sg_sum[subgroup_id] = chunk_sum;
|
|
212
|
+
}
|
|
213
|
+
workgroupBarrier();
|
|
214
|
+
|
|
215
|
+
if (tid == 0u) {
|
|
216
|
+
var s: f32 = 0.0;
|
|
217
|
+
for (var i: u32 = 0u; i < num_subgroups; i++) {
|
|
218
|
+
s = s + sg_sum[i];
|
|
219
|
+
}
|
|
220
|
+
global_sum = s;
|
|
221
|
+
}
|
|
222
|
+
workgroupBarrier();
|
|
223
|
+
|
|
224
|
+
running_sum = running_sum * rescale + global_sum;
|
|
225
|
+
running_max = new_max;
|
|
226
|
+
|
|
227
|
+
out_accum = out_accum * rescale;
|
|
228
|
+
let chunk_len = min(WORKGROUP_SIZE, kv_len - k_start);
|
|
229
|
+
let chunk_len4 = (chunk_len / 4u) * 4u;
|
|
230
|
+
if (USE_DIRECT_KV_LAYOUT) {
|
|
231
|
+
let direct_v_base = (u.kv_start + k_start) * u.num_kv_heads * HEAD_DIM + kv_head_idx * HEAD_DIM;
|
|
232
|
+
let direct_v_stride = u.num_kv_heads * HEAD_DIM;
|
|
233
|
+
for (var k: u32 = 0u; k < chunk_len4; k = k + 4u) {
|
|
234
|
+
let v_base0 = direct_v_base + k * direct_v_stride;
|
|
235
|
+
let v_base1 = direct_v_base + (k + 1u) * direct_v_stride;
|
|
236
|
+
let v_base2 = direct_v_base + (k + 2u) * direct_v_stride;
|
|
237
|
+
let v_base3 = direct_v_base + (k + 3u) * direct_v_stride;
|
|
238
|
+
out_accum = out_accum + shared_scores[k] * f32(V[v_base0 + tid]);
|
|
239
|
+
out_accum = out_accum + shared_scores[k + 1u] * f32(V[v_base1 + tid]);
|
|
240
|
+
out_accum = out_accum + shared_scores[k + 2u] * f32(V[v_base2 + tid]);
|
|
241
|
+
out_accum = out_accum + shared_scores[k + 3u] * f32(V[v_base3 + tid]);
|
|
242
|
+
}
|
|
243
|
+
for (var k: u32 = chunk_len4; k < chunk_len; k = k + 1u) {
|
|
244
|
+
let v_base = direct_v_base + k * direct_v_stride;
|
|
245
|
+
out_accum = out_accum + shared_scores[k] * f32(V[v_base + tid]);
|
|
246
|
+
}
|
|
247
|
+
} else {
|
|
248
|
+
for (var k: u32 = 0u; k < chunk_len4; k = k + 4u) {
|
|
249
|
+
let v_base0 = shared_v_bases[k];
|
|
250
|
+
let v_base1 = shared_v_bases[k + 1u];
|
|
251
|
+
let v_base2 = shared_v_bases[k + 2u];
|
|
252
|
+
let v_base3 = shared_v_bases[k + 3u];
|
|
253
|
+
out_accum = out_accum + shared_scores[k] * f32(V[v_base0 + tid]);
|
|
254
|
+
out_accum = out_accum + shared_scores[k + 1u] * f32(V[v_base1 + tid]);
|
|
255
|
+
out_accum = out_accum + shared_scores[k + 2u] * f32(V[v_base2 + tid]);
|
|
256
|
+
out_accum = out_accum + shared_scores[k + 3u] * f32(V[v_base3 + tid]);
|
|
257
|
+
}
|
|
258
|
+
for (var k: u32 = chunk_len4; k < chunk_len; k = k + 1u) {
|
|
259
|
+
let v_base = shared_v_bases[k];
|
|
260
|
+
out_accum = out_accum + shared_scores[k] * f32(V[v_base + tid]);
|
|
261
|
+
}
|
|
262
|
+
}
|
|
263
|
+
workgroupBarrier();
|
|
264
|
+
}
|
|
265
|
+
|
|
266
|
+
let inv_sum = select(0.0, 1.0 / running_sum, running_sum > 0.0);
|
|
267
|
+
output[q_offset + tid] = out_accum * inv_sum * sigmoid(output_gate[q_offset + tid]);
|
|
268
|
+
}
|
|
@@ -0,0 +1,269 @@
|
|
|
1
|
+
// Fused Multi-Head Attention Kernel (head_dim 128, f16 KV)
|
|
2
|
+
//
|
|
3
|
+
// Fixed-shape prefill kernel for 128-dim heads. It reuses one shared 128-dim
|
|
4
|
+
// block for K and then V so the workgroup avoids the repeated 32-wide tiling
|
|
5
|
+
// used by the generic small-head path.
|
|
6
|
+
|
|
7
|
+
enable f16;
|
|
8
|
+
|
|
9
|
+
const BLOCK_SIZE: u32 = 32u;
|
|
10
|
+
const WORKGROUP_SIZE: u32 = BLOCK_SIZE;
|
|
11
|
+
const HEAD_DIM: u32 = 128u;
|
|
12
|
+
const HEAD_DIM_VECS: u32 = 32u;
|
|
13
|
+
|
|
14
|
+
struct Uniforms {
|
|
15
|
+
num_heads: u32,
|
|
16
|
+
num_kv_heads: u32,
|
|
17
|
+
head_dim: u32,
|
|
18
|
+
seq_len: u32,
|
|
19
|
+
query_len: u32,
|
|
20
|
+
scale: f32,
|
|
21
|
+
is_causal: u32,
|
|
22
|
+
start_pos: u32,
|
|
23
|
+
attn_softcap: f32,
|
|
24
|
+
sliding_window: u32,
|
|
25
|
+
kv_len_source: u32,
|
|
26
|
+
kv_start: u32,
|
|
27
|
+
page_size: u32,
|
|
28
|
+
kv_layout: u32,
|
|
29
|
+
_pad: u32,
|
|
30
|
+
}
|
|
31
|
+
|
|
32
|
+
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
33
|
+
@group(0) @binding(1) var<storage, read> Q: array<f32>;
|
|
34
|
+
@group(0) @binding(2) var<storage, read> K: array<f16>;
|
|
35
|
+
@group(0) @binding(3) var<storage, read> V: array<f16>;
|
|
36
|
+
@group(0) @binding(4) var<storage, read_write> output: array<f32>;
|
|
37
|
+
@group(0) @binding(5) var<storage, read> kv_len_buffer: array<u32>;
|
|
38
|
+
@group(0) @binding(6) var<storage, read> page_table: array<u32>;
|
|
39
|
+
|
|
40
|
+
var<workgroup> shared_block: array<vec4<f16>, BLOCK_SIZE * HEAD_DIM_VECS>;
|
|
41
|
+
|
|
42
|
+
fn zero_vec4_f16() -> vec4<f16> {
|
|
43
|
+
return vec4<f16>(f16(0.0), f16(0.0), f16(0.0), f16(0.0));
|
|
44
|
+
}
|
|
45
|
+
|
|
46
|
+
fn get_kv_head_idx(query_head_idx: u32) -> u32 {
|
|
47
|
+
let heads_per_kv = u.num_heads / u.num_kv_heads;
|
|
48
|
+
return query_head_idx / heads_per_kv;
|
|
49
|
+
}
|
|
50
|
+
|
|
51
|
+
fn get_kv_pos(key_pos: u32) -> u32 {
|
|
52
|
+
let abs_key = u.kv_start + key_pos;
|
|
53
|
+
if (u.kv_layout == 1u && u.sliding_window > 0u) {
|
|
54
|
+
return abs_key % u.sliding_window;
|
|
55
|
+
}
|
|
56
|
+
if (u.kv_layout == 2u) {
|
|
57
|
+
let page_idx = abs_key / u.page_size;
|
|
58
|
+
let in_page = abs_key - (page_idx * u.page_size);
|
|
59
|
+
let phys_page = page_table[page_idx];
|
|
60
|
+
return phys_page * u.page_size + in_page;
|
|
61
|
+
}
|
|
62
|
+
return abs_key;
|
|
63
|
+
}
|
|
64
|
+
|
|
65
|
+
fn get_kv_len() -> u32 {
|
|
66
|
+
if (u.kv_len_source == 0u) {
|
|
67
|
+
return u.seq_len;
|
|
68
|
+
}
|
|
69
|
+
return kv_len_buffer[0];
|
|
70
|
+
}
|
|
71
|
+
|
|
72
|
+
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
73
|
+
fn main(
|
|
74
|
+
@builtin(local_invocation_id) local_id: vec3<u32>,
|
|
75
|
+
@builtin(workgroup_id) wg_id: vec3<u32>
|
|
76
|
+
) {
|
|
77
|
+
if (u.head_dim != HEAD_DIM) {
|
|
78
|
+
return;
|
|
79
|
+
}
|
|
80
|
+
|
|
81
|
+
let linear = wg_id.x;
|
|
82
|
+
let num_heads = u.num_heads;
|
|
83
|
+
let head_idx = linear % num_heads;
|
|
84
|
+
let query_block_idx = linear / num_heads;
|
|
85
|
+
let thread_idx = local_id.x;
|
|
86
|
+
|
|
87
|
+
let kv_head_idx = get_kv_head_idx(head_idx);
|
|
88
|
+
let seq_len = get_kv_len();
|
|
89
|
+
let query_len = u.query_len;
|
|
90
|
+
let scale = u.scale;
|
|
91
|
+
|
|
92
|
+
let query_pos = query_block_idx * BLOCK_SIZE + thread_idx;
|
|
93
|
+
let valid_query = query_pos < query_len;
|
|
94
|
+
let abs_query = query_pos + u.start_pos;
|
|
95
|
+
|
|
96
|
+
var q_local: array<vec4<f32>, HEAD_DIM_VECS>;
|
|
97
|
+
var acc: array<vec4<f32>, HEAD_DIM_VECS>;
|
|
98
|
+
|
|
99
|
+
if (valid_query) {
|
|
100
|
+
let q_offset = query_pos * num_heads * HEAD_DIM + head_idx * HEAD_DIM;
|
|
101
|
+
for (var d4: u32 = 0u; d4 < HEAD_DIM_VECS; d4 = d4 + 1u) {
|
|
102
|
+
let base = q_offset + d4 * 4u;
|
|
103
|
+
q_local[d4] = vec4<f32>(
|
|
104
|
+
Q[base],
|
|
105
|
+
Q[base + 1u],
|
|
106
|
+
Q[base + 2u],
|
|
107
|
+
Q[base + 3u]
|
|
108
|
+
);
|
|
109
|
+
acc[d4] = vec4<f32>(0.0);
|
|
110
|
+
}
|
|
111
|
+
}
|
|
112
|
+
|
|
113
|
+
var m_i: f32 = -3.402823e+38;
|
|
114
|
+
var l_i: f32 = 0.0;
|
|
115
|
+
|
|
116
|
+
let num_kv_blocks = (seq_len + BLOCK_SIZE - 1u) / BLOCK_SIZE;
|
|
117
|
+
var min_key_pos: u32 = 0u;
|
|
118
|
+
var max_key_pos: u32 = seq_len;
|
|
119
|
+
|
|
120
|
+
if (valid_query) {
|
|
121
|
+
if (u.is_causal != 0u) {
|
|
122
|
+
if (abs_query < u.kv_start) {
|
|
123
|
+
max_key_pos = 0u;
|
|
124
|
+
} else {
|
|
125
|
+
let causal_limit = abs_query - u.kv_start + 1u;
|
|
126
|
+
max_key_pos = min(seq_len, causal_limit);
|
|
127
|
+
}
|
|
128
|
+
}
|
|
129
|
+
if (u.sliding_window > 0u && abs_query >= u.sliding_window) {
|
|
130
|
+
let min_abs_key = abs_query - u.sliding_window + 1u;
|
|
131
|
+
if (min_abs_key > u.kv_start) {
|
|
132
|
+
min_key_pos = min_abs_key - u.kv_start;
|
|
133
|
+
}
|
|
134
|
+
min_key_pos = min(min_key_pos, seq_len);
|
|
135
|
+
}
|
|
136
|
+
}
|
|
137
|
+
|
|
138
|
+
for (var kv_block: u32 = 0u; kv_block < num_kv_blocks; kv_block = kv_block + 1u) {
|
|
139
|
+
let kv_block_start = kv_block * BLOCK_SIZE;
|
|
140
|
+
|
|
141
|
+
var scores: array<f32, BLOCK_SIZE>;
|
|
142
|
+
var key_active: array<u32, BLOCK_SIZE>;
|
|
143
|
+
var probs: array<f32, BLOCK_SIZE>;
|
|
144
|
+
for (var k_init: u32 = 0u; k_init < BLOCK_SIZE; k_init = k_init + 1u) {
|
|
145
|
+
scores[k_init] = 0.0;
|
|
146
|
+
key_active[k_init] = 0u;
|
|
147
|
+
probs[k_init] = 0.0;
|
|
148
|
+
}
|
|
149
|
+
|
|
150
|
+
if (valid_query) {
|
|
151
|
+
for (var k_mask: u32 = 0u; k_mask < BLOCK_SIZE; k_mask = k_mask + 1u) {
|
|
152
|
+
let key_pos = kv_block_start + k_mask;
|
|
153
|
+
if (key_pos < seq_len && key_pos >= min_key_pos && key_pos < max_key_pos) {
|
|
154
|
+
key_active[k_mask] = 1u;
|
|
155
|
+
}
|
|
156
|
+
}
|
|
157
|
+
}
|
|
158
|
+
|
|
159
|
+
let key_pos_load = kv_block_start + thread_idx;
|
|
160
|
+
let shared_row = thread_idx * HEAD_DIM_VECS;
|
|
161
|
+
if (key_pos_load < seq_len) {
|
|
162
|
+
let k_idx = get_kv_pos(key_pos_load);
|
|
163
|
+
let k_offset = k_idx * u.num_kv_heads * HEAD_DIM + kv_head_idx * HEAD_DIM;
|
|
164
|
+
for (var d4: u32 = 0u; d4 < HEAD_DIM_VECS; d4 = d4 + 1u) {
|
|
165
|
+
let base = k_offset + d4 * 4u;
|
|
166
|
+
shared_block[shared_row + d4] = vec4<f16>(
|
|
167
|
+
K[base],
|
|
168
|
+
K[base + 1u],
|
|
169
|
+
K[base + 2u],
|
|
170
|
+
K[base + 3u]
|
|
171
|
+
);
|
|
172
|
+
}
|
|
173
|
+
} else {
|
|
174
|
+
for (var d4: u32 = 0u; d4 < HEAD_DIM_VECS; d4 = d4 + 1u) {
|
|
175
|
+
shared_block[shared_row + d4] = zero_vec4_f16();
|
|
176
|
+
}
|
|
177
|
+
}
|
|
178
|
+
|
|
179
|
+
workgroupBarrier();
|
|
180
|
+
|
|
181
|
+
var m_new: f32 = m_i;
|
|
182
|
+
if (valid_query) {
|
|
183
|
+
var block_max: f32 = -3.402823e+38;
|
|
184
|
+
for (var k: u32 = 0u; k < BLOCK_SIZE; k = k + 1u) {
|
|
185
|
+
if (key_active[k] == 0u) { continue; }
|
|
186
|
+
|
|
187
|
+
let key_row = k * HEAD_DIM_VECS;
|
|
188
|
+
var dot_partial: f32 = 0.0;
|
|
189
|
+
for (var d4: u32 = 0u; d4 < HEAD_DIM_VECS; d4 = d4 + 1u) {
|
|
190
|
+
dot_partial = dot_partial + dot(
|
|
191
|
+
q_local[d4],
|
|
192
|
+
vec4<f32>(shared_block[key_row + d4])
|
|
193
|
+
);
|
|
194
|
+
}
|
|
195
|
+
|
|
196
|
+
var s = dot_partial * scale;
|
|
197
|
+
if (u.attn_softcap > 0.0) {
|
|
198
|
+
s = tanh(s / u.attn_softcap) * u.attn_softcap;
|
|
199
|
+
}
|
|
200
|
+
scores[k] = s;
|
|
201
|
+
block_max = max(block_max, s);
|
|
202
|
+
}
|
|
203
|
+
|
|
204
|
+
m_new = max(m_i, block_max);
|
|
205
|
+
let correction = exp(m_i - m_new);
|
|
206
|
+
l_i = l_i * correction;
|
|
207
|
+
for (var d4: u32 = 0u; d4 < HEAD_DIM_VECS; d4 = d4 + 1u) {
|
|
208
|
+
acc[d4] = acc[d4] * correction;
|
|
209
|
+
}
|
|
210
|
+
|
|
211
|
+
for (var k: u32 = 0u; k < BLOCK_SIZE; k = k + 1u) {
|
|
212
|
+
if (key_active[k] == 0u) { continue; }
|
|
213
|
+
let p = exp(scores[k] - m_new);
|
|
214
|
+
probs[k] = p;
|
|
215
|
+
l_i = l_i + p;
|
|
216
|
+
}
|
|
217
|
+
}
|
|
218
|
+
|
|
219
|
+
workgroupBarrier();
|
|
220
|
+
|
|
221
|
+
if (key_pos_load < seq_len) {
|
|
222
|
+
let v_idx = get_kv_pos(key_pos_load);
|
|
223
|
+
let v_offset = v_idx * u.num_kv_heads * HEAD_DIM + kv_head_idx * HEAD_DIM;
|
|
224
|
+
for (var d4: u32 = 0u; d4 < HEAD_DIM_VECS; d4 = d4 + 1u) {
|
|
225
|
+
let base = v_offset + d4 * 4u;
|
|
226
|
+
shared_block[shared_row + d4] = vec4<f16>(
|
|
227
|
+
V[base],
|
|
228
|
+
V[base + 1u],
|
|
229
|
+
V[base + 2u],
|
|
230
|
+
V[base + 3u]
|
|
231
|
+
);
|
|
232
|
+
}
|
|
233
|
+
} else {
|
|
234
|
+
for (var d4: u32 = 0u; d4 < HEAD_DIM_VECS; d4 = d4 + 1u) {
|
|
235
|
+
shared_block[shared_row + d4] = zero_vec4_f16();
|
|
236
|
+
}
|
|
237
|
+
}
|
|
238
|
+
|
|
239
|
+
workgroupBarrier();
|
|
240
|
+
|
|
241
|
+
if (valid_query) {
|
|
242
|
+
for (var k: u32 = 0u; k < BLOCK_SIZE; k = k + 1u) {
|
|
243
|
+
let p = probs[k];
|
|
244
|
+
if (p == 0.0) { continue; }
|
|
245
|
+
|
|
246
|
+
let value_row = k * HEAD_DIM_VECS;
|
|
247
|
+
for (var d4: u32 = 0u; d4 < HEAD_DIM_VECS; d4 = d4 + 1u) {
|
|
248
|
+
acc[d4] = acc[d4] + p * vec4<f32>(shared_block[value_row + d4]);
|
|
249
|
+
}
|
|
250
|
+
}
|
|
251
|
+
m_i = m_new;
|
|
252
|
+
}
|
|
253
|
+
|
|
254
|
+
workgroupBarrier();
|
|
255
|
+
}
|
|
256
|
+
|
|
257
|
+
if (valid_query) {
|
|
258
|
+
let out_offset = query_pos * num_heads * HEAD_DIM + head_idx * HEAD_DIM;
|
|
259
|
+
let inv_l_i = select(0.0, 1.0 / l_i, l_i > 0.0);
|
|
260
|
+
for (var d4: u32 = 0u; d4 < HEAD_DIM_VECS; d4 = d4 + 1u) {
|
|
261
|
+
let out_vec = acc[d4] * inv_l_i;
|
|
262
|
+
let base = out_offset + d4 * 4u;
|
|
263
|
+
output[base] = out_vec.x;
|
|
264
|
+
output[base + 1u] = out_vec.y;
|
|
265
|
+
output[base + 2u] = out_vec.z;
|
|
266
|
+
output[base + 3u] = out_vec.w;
|
|
267
|
+
}
|
|
268
|
+
}
|
|
269
|
+
}
|