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
|
@@ -1,24 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* Attention Module - Re-exports
|
|
3
|
-
*
|
|
4
|
-
* @module inference/pipelines/text/attention
|
|
5
|
-
*/
|
|
6
|
-
|
|
7
|
-
// Types and utilities
|
|
8
|
-
export {
|
|
9
|
-
type AttentionConfig,
|
|
10
|
-
type AttentionState,
|
|
11
|
-
type AttentionResult,
|
|
12
|
-
type AttentionDebugFlags,
|
|
13
|
-
shouldDebugLayer,
|
|
14
|
-
markStageLogged,
|
|
15
|
-
releaseOrTrack,
|
|
16
|
-
getQKNormOnesBuffer,
|
|
17
|
-
} from './types.js';
|
|
18
|
-
|
|
19
|
-
// Run (immediate submission)
|
|
20
|
-
export { runLayerAttentionGPU } from './run.js';
|
|
21
|
-
|
|
22
|
-
// Record (batched submission)
|
|
23
|
-
export { recordLayerAttentionGPU } from './record.js';
|
|
24
|
-
export { runAttentionBDPA, recordAttentionBDPA } from '../../../../gpu/kernel-selector.js';
|
|
@@ -1,12 +0,0 @@
|
|
|
1
|
-
import type { Tensor } from '../../../../gpu/tensor.js';
|
|
2
|
-
|
|
3
|
-
export interface AttentionProjectionInputResult {
|
|
4
|
-
oProjInput: Tensor;
|
|
5
|
-
oProjInputTemp: Tensor | null;
|
|
6
|
-
}
|
|
7
|
-
|
|
8
|
-
export function prepareAttentionProjectionInput(
|
|
9
|
-
attnForProjection: Tensor,
|
|
10
|
-
matmulOutputDtype: string,
|
|
11
|
-
castToF16: (tensor: Tensor) => Promise<Tensor>
|
|
12
|
-
): Promise<AttentionProjectionInputResult>;
|
|
@@ -1,36 +0,0 @@
|
|
|
1
|
-
export interface AttentionPrecisionContract {
|
|
2
|
-
precision: {
|
|
3
|
-
activationDtype?: 'f16' | 'f32';
|
|
4
|
-
kvDtype?: 'f16' | 'f32';
|
|
5
|
-
inputDtype?: 'f16' | 'f32';
|
|
6
|
-
outputDtype?: 'f16' | 'f32';
|
|
7
|
-
} | null;
|
|
8
|
-
explicitInputDtype: 'f16' | 'f32' | null;
|
|
9
|
-
explicitOutputDtype: 'f16' | 'f32' | null;
|
|
10
|
-
explicitKvDtype: 'f16' | 'f32' | null;
|
|
11
|
-
resolvedActivationDtype: 'f16' | 'f32' | null;
|
|
12
|
-
resolvedOutputDtype: 'f16' | 'f32' | null;
|
|
13
|
-
resolvedKvCacheDtype: 'f16' | 'f32' | null;
|
|
14
|
-
}
|
|
15
|
-
|
|
16
|
-
export function resolveAttentionPrecisionContract(
|
|
17
|
-
config: {
|
|
18
|
-
isPrefill?: boolean;
|
|
19
|
-
layerIdx?: number;
|
|
20
|
-
kernelPath?: Record<string, unknown> | null;
|
|
21
|
-
activationDtype?: 'f16' | 'f32' | null;
|
|
22
|
-
inputDtype?: 'f16' | 'f32' | null;
|
|
23
|
-
outputDtype?: 'f16' | 'f32' | null;
|
|
24
|
-
kvDtype?: 'f16' | 'f32' | null;
|
|
25
|
-
} | null | undefined,
|
|
26
|
-
state: {
|
|
27
|
-
kvCache?: {
|
|
28
|
-
kvDtype?: 'f16' | 'f32' | null;
|
|
29
|
-
} | null;
|
|
30
|
-
} | null | undefined
|
|
31
|
-
): AttentionPrecisionContract;
|
|
32
|
-
|
|
33
|
-
export function isAttentionKvDtypeExplicit(
|
|
34
|
-
contract: AttentionPrecisionContract | null | undefined,
|
|
35
|
-
targetDtype: 'f16' | 'f32'
|
|
36
|
-
): boolean;
|
|
@@ -1,208 +0,0 @@
|
|
|
1
|
-
import type { CommandRecorder } from '../../../../gpu/kernel-selector.js';
|
|
2
|
-
import type { Tensor } from '../../../../gpu/tensor.js';
|
|
3
|
-
import type { WeightBuffer, CpuWeightBuffer } from '../../../../gpu/weight-buffer.js';
|
|
4
|
-
import type { LayerWeights } from '../types.js';
|
|
5
|
-
import type { LoRAAdapter } from '../lora.js';
|
|
6
|
-
import type { MatmulDebugConfigSchema } from '../../../../config/schema/debug.schema.js';
|
|
7
|
-
|
|
8
|
-
export interface AttentionInputInfo {
|
|
9
|
-
phase: 'prefill' | 'decode';
|
|
10
|
-
layerIdx: number;
|
|
11
|
-
numTokens?: number;
|
|
12
|
-
kvLen?: number;
|
|
13
|
-
numHeads?: number;
|
|
14
|
-
numKVHeads?: number;
|
|
15
|
-
headDim?: number;
|
|
16
|
-
activationDtype?: string | null;
|
|
17
|
-
inputDtype?: string | null;
|
|
18
|
-
normedDtype?: string | null;
|
|
19
|
-
kvDtype?: string | null;
|
|
20
|
-
kvCacheDtype?: string | null;
|
|
21
|
-
cachedKDtype?: string | null;
|
|
22
|
-
cachedVDtype?: string | null;
|
|
23
|
-
qDtype?: string | null;
|
|
24
|
-
kDtype?: string | null;
|
|
25
|
-
vDtype?: string | null;
|
|
26
|
-
qWeightDtype?: string | null;
|
|
27
|
-
kWeightDtype?: string | null;
|
|
28
|
-
vWeightDtype?: string | null;
|
|
29
|
-
oWeightDtype?: string | null;
|
|
30
|
-
useF16Attention?: boolean;
|
|
31
|
-
useF16Activations?: boolean;
|
|
32
|
-
hasF16Weights?: boolean;
|
|
33
|
-
matmulOutputDtype?: string | null;
|
|
34
|
-
useFusedQKV?: boolean;
|
|
35
|
-
kvStart?: number;
|
|
36
|
-
kvLayout?: string;
|
|
37
|
-
kvPageSize?: number | null;
|
|
38
|
-
hotLen?: number | null;
|
|
39
|
-
coldLen?: number | null;
|
|
40
|
-
hotWindow?: number | null;
|
|
41
|
-
hotStart?: number | null;
|
|
42
|
-
coldQuantMode?: string | null;
|
|
43
|
-
}
|
|
44
|
-
|
|
45
|
-
export function recordAttentionInputs(
|
|
46
|
-
state: { stats?: { attentionInputs?: AttentionInputInfo[] } } | null | undefined,
|
|
47
|
-
info: AttentionInputInfo | null | undefined
|
|
48
|
-
): void;
|
|
49
|
-
|
|
50
|
-
export function shouldForceF32AttentionProjectionForRoPE(options: {
|
|
51
|
-
attentionInputDtype: string;
|
|
52
|
-
headDim: number;
|
|
53
|
-
rotaryDim?: number;
|
|
54
|
-
interleaved?: boolean;
|
|
55
|
-
kernelPathIsF16?: boolean;
|
|
56
|
-
}): boolean;
|
|
57
|
-
export function resolveAttentionProjectionOutputDtype(
|
|
58
|
-
attentionInputDtype: string,
|
|
59
|
-
options?: { forceF32?: boolean }
|
|
60
|
-
): 'f16' | 'f32' | string;
|
|
61
|
-
export function resolveProjectionMatmulDtype(options: {
|
|
62
|
-
useFusedQKV: boolean;
|
|
63
|
-
phase: 'prefill' | 'decode';
|
|
64
|
-
layerIdx: number;
|
|
65
|
-
kernelPath: Record<string, unknown> | null | undefined;
|
|
66
|
-
precisionField: 'inputDtype' | 'outputDtype';
|
|
67
|
-
fallbackDtype: 'f16' | 'f32' | string | null | undefined;
|
|
68
|
-
}): 'f16' | 'f32' | string | null | undefined;
|
|
69
|
-
export function resolveProjectionSliceOffsetBytes(
|
|
70
|
-
weightBuffer: WeightBuffer | Tensor | GPUBuffer | null | undefined,
|
|
71
|
-
outputRows: number,
|
|
72
|
-
inputCols: number
|
|
73
|
-
): number;
|
|
74
|
-
|
|
75
|
-
export interface AttentionQKNormState {
|
|
76
|
-
wantsQKNorm: boolean;
|
|
77
|
-
hasQNorm: boolean;
|
|
78
|
-
hasKNorm: boolean;
|
|
79
|
-
allowUnitQKNorm: boolean;
|
|
80
|
-
skipKNorm: boolean;
|
|
81
|
-
}
|
|
82
|
-
|
|
83
|
-
export function hasAttentionProjectionDiagnostics(state: {
|
|
84
|
-
operatorDiagnostics?: {
|
|
85
|
-
enabled?: boolean;
|
|
86
|
-
tsirFixture?: { dir?: string | null } | null;
|
|
87
|
-
} | null;
|
|
88
|
-
debugProbes?: Array<{ stage?: string | null }> | null;
|
|
89
|
-
} | null | undefined): boolean;
|
|
90
|
-
|
|
91
|
-
export function hasAttentionStageDiagnostics(
|
|
92
|
-
state: {
|
|
93
|
-
operatorDiagnostics?: {
|
|
94
|
-
enabled?: boolean;
|
|
95
|
-
tsirFixture?: { dir?: string | null } | null;
|
|
96
|
-
} | null;
|
|
97
|
-
debugProbes?: Array<{ stage?: string | null }> | null;
|
|
98
|
-
} | null | undefined,
|
|
99
|
-
stages: string[]
|
|
100
|
-
): boolean;
|
|
101
|
-
|
|
102
|
-
export function resolveAttentionQKNormState(options: {
|
|
103
|
-
config: {
|
|
104
|
-
queryKeyNorm?: boolean;
|
|
105
|
-
queryKeyNormWeightLayers?: number[] | null;
|
|
106
|
-
};
|
|
107
|
-
layerWeights: LayerWeights;
|
|
108
|
-
layerIdx: number;
|
|
109
|
-
reusesSharedKV: boolean;
|
|
110
|
-
}): AttentionQKNormState;
|
|
111
|
-
|
|
112
|
-
export interface ProjectAttentionQKNormFusionOptions {
|
|
113
|
-
enabled: boolean;
|
|
114
|
-
getNormWeightBuffer?: (weight: GPUBuffer | Float32Array | ArrayBuffer | CpuWeightBuffer, label: string) => GPUBuffer;
|
|
115
|
-
rmsNormEps: number;
|
|
116
|
-
rmsNormWeightOffset?: boolean;
|
|
117
|
-
skipKNorm?: boolean;
|
|
118
|
-
allowUnitQKNorm?: boolean;
|
|
119
|
-
projectionDiagnosticsEnabled?: boolean;
|
|
120
|
-
}
|
|
121
|
-
|
|
122
|
-
export interface ProjectAttentionQKNormRoPEFusionOptions extends ProjectAttentionQKNormFusionOptions {
|
|
123
|
-
freqsCos?: GPUBuffer | Tensor | null;
|
|
124
|
-
freqsSin?: GPUBuffer | Tensor | null;
|
|
125
|
-
headDim?: number;
|
|
126
|
-
startPos?: number;
|
|
127
|
-
rotaryDim?: number;
|
|
128
|
-
pairSpanDim?: number;
|
|
129
|
-
interleaved?: boolean;
|
|
130
|
-
reusesSharedKV?: boolean;
|
|
131
|
-
f16KVCacheWrite?: {
|
|
132
|
-
keysBuffer: GPUBuffer;
|
|
133
|
-
valuesBuffer: GPUBuffer;
|
|
134
|
-
dstOffset: number;
|
|
135
|
-
} | null;
|
|
136
|
-
}
|
|
137
|
-
|
|
138
|
-
export interface ProjectAttentionQKVOptions {
|
|
139
|
-
recorder?: CommandRecorder | null;
|
|
140
|
-
normed: Tensor;
|
|
141
|
-
layerWeights: LayerWeights;
|
|
142
|
-
numTokens: number;
|
|
143
|
-
numHeads: number;
|
|
144
|
-
numKVHeads: number;
|
|
145
|
-
headDim: number;
|
|
146
|
-
hiddenSize: number;
|
|
147
|
-
layerIdx: number;
|
|
148
|
-
matmulOutputDtype: string;
|
|
149
|
-
getWeightBuffer?: (weight: GPUBuffer | WeightBuffer | Float32Array | ArrayBuffer | CpuWeightBuffer, label: string) => GPUBuffer | WeightBuffer;
|
|
150
|
-
lora?: LoRAAdapter | null;
|
|
151
|
-
releaseTemporary: (buffer: GPUBuffer) => void;
|
|
152
|
-
matmulDebug?: MatmulDebugConfigSchema | null;
|
|
153
|
-
onFusedQKV?: ((info: { qSize: number; kSize: number; vSize: number; totalSize: number }) => void) | null;
|
|
154
|
-
qkNormFusion?: ProjectAttentionQKNormFusionOptions | null;
|
|
155
|
-
qkNormRoPEFusion?: ProjectAttentionQKNormRoPEFusionOptions | null;
|
|
156
|
-
}
|
|
157
|
-
|
|
158
|
-
export interface ProjectAttentionQKVResult {
|
|
159
|
-
qTensor: Tensor;
|
|
160
|
-
qGateTensor: Tensor | null;
|
|
161
|
-
kTensor: Tensor | null;
|
|
162
|
-
vTensor: Tensor | null;
|
|
163
|
-
usedFusedQKV: boolean;
|
|
164
|
-
valueAliasesKey: boolean;
|
|
165
|
-
qkNormApplied: boolean;
|
|
166
|
-
ropeApplied: boolean;
|
|
167
|
-
kvCacheWriteFused: boolean;
|
|
168
|
-
}
|
|
169
|
-
|
|
170
|
-
export function projectAttentionQKV(options: ProjectAttentionQKVOptions): Promise<ProjectAttentionQKVResult>;
|
|
171
|
-
|
|
172
|
-
export interface ApplyAttentionQKNormOptions {
|
|
173
|
-
recorder?: CommandRecorder | null;
|
|
174
|
-
qTensor: Tensor;
|
|
175
|
-
kTensor: Tensor;
|
|
176
|
-
layerWeights: LayerWeights;
|
|
177
|
-
getNormWeightBuffer?: (weight: GPUBuffer | Float32Array | ArrayBuffer | CpuWeightBuffer, label: string) => GPUBuffer;
|
|
178
|
-
rmsNormEps: number;
|
|
179
|
-
numTokens: number;
|
|
180
|
-
numHeads: number;
|
|
181
|
-
numKVHeads: number;
|
|
182
|
-
headDim: number;
|
|
183
|
-
rmsNormWeightOffset?: boolean;
|
|
184
|
-
releaseTemporary: (buffer: GPUBuffer) => void;
|
|
185
|
-
onQNormApplied?: ((tensor: Tensor) => Promise<void> | void) | null;
|
|
186
|
-
onKNormApplied?: ((tensor: Tensor) => Promise<void> | void) | null;
|
|
187
|
-
retainKInput?: boolean;
|
|
188
|
-
allowUnitQKNorm?: boolean;
|
|
189
|
-
}
|
|
190
|
-
|
|
191
|
-
export function applyAttentionQKNorm(
|
|
192
|
-
options: ApplyAttentionQKNormOptions
|
|
193
|
-
): Promise<{ qTensor: Tensor; kTensor: Tensor }>;
|
|
194
|
-
|
|
195
|
-
export interface ApplyAttentionValueNormOptions {
|
|
196
|
-
recorder?: CommandRecorder | null;
|
|
197
|
-
vTensor: Tensor;
|
|
198
|
-
rmsNormEps: number;
|
|
199
|
-
numTokens: number;
|
|
200
|
-
numKVHeads: number;
|
|
201
|
-
headDim: number;
|
|
202
|
-
releaseTemporary: (buffer: GPUBuffer) => void;
|
|
203
|
-
onVNormApplied?: ((tensor: Tensor) => Promise<void> | void) | null;
|
|
204
|
-
}
|
|
205
|
-
|
|
206
|
-
export function applyAttentionValueNorm(
|
|
207
|
-
options: ApplyAttentionValueNormOptions
|
|
208
|
-
): Promise<Tensor>;
|
|
@@ -1,38 +0,0 @@
|
|
|
1
|
-
export interface QuantAttentionRuntimeState {
|
|
2
|
-
coldLen?: number;
|
|
3
|
-
hotLen?: number;
|
|
4
|
-
hotWindow?: number;
|
|
5
|
-
hotStart?: number;
|
|
6
|
-
coldPackedStride?: number;
|
|
7
|
-
coldQuantMode?: string;
|
|
8
|
-
prodMode?: boolean;
|
|
9
|
-
rotationMatrixBuffer?: GPUBuffer | null;
|
|
10
|
-
codebookCentroidsBuffer?: GPUBuffer | null;
|
|
11
|
-
residualKGPU?: GPUBuffer | null;
|
|
12
|
-
residualVGPU?: GPUBuffer | null;
|
|
13
|
-
residualNormsKGPU?: GPUBuffer | null;
|
|
14
|
-
residualNormsVGPU?: GPUBuffer | null;
|
|
15
|
-
qjlMatrixBuffer?: GPUBuffer | null;
|
|
16
|
-
residualPackedStride?: number;
|
|
17
|
-
}
|
|
18
|
-
|
|
19
|
-
export interface QuantAttentionInputOptions {
|
|
20
|
-
seqLen?: number;
|
|
21
|
-
kvLen?: number;
|
|
22
|
-
numKVHeads?: number;
|
|
23
|
-
causal?: boolean;
|
|
24
|
-
startPos?: number;
|
|
25
|
-
slidingWindow?: number;
|
|
26
|
-
attnSoftcap?: number;
|
|
27
|
-
scale?: number;
|
|
28
|
-
}
|
|
29
|
-
|
|
30
|
-
export declare function buildTieredQuantAttentionOptions(
|
|
31
|
-
kvState: QuantAttentionRuntimeState | null | undefined,
|
|
32
|
-
options?: QuantAttentionInputOptions
|
|
33
|
-
): Record<string, unknown>;
|
|
34
|
-
|
|
35
|
-
export declare function buildContiguousQuantAttentionOptions(
|
|
36
|
-
kvState: QuantAttentionRuntimeState | null | undefined,
|
|
37
|
-
options?: QuantAttentionInputOptions
|
|
38
|
-
): Record<string, unknown>;
|
|
@@ -1,36 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* Attention Record - Batched GPU recording path
|
|
3
|
-
*
|
|
4
|
-
* Contains recordLayerAttentionGPU which records attention operations
|
|
5
|
-
* to a shared command encoder without submitting. All operations are
|
|
6
|
-
* batched and submitted together at the end of the forward pass.
|
|
7
|
-
*
|
|
8
|
-
* @module inference/pipelines/text/attention/record
|
|
9
|
-
*/
|
|
10
|
-
|
|
11
|
-
import type { LayerWeights } from '../types.js';
|
|
12
|
-
import type { WeightBuffer, CpuWeightBuffer } from '../../../../gpu/weight-buffer.js';
|
|
13
|
-
import type { CommandRecorder } from '../../../../gpu/kernel-selector.js';
|
|
14
|
-
import type { Tensor } from '../../../../gpu/tensor.js';
|
|
15
|
-
import type { LoRAAdapter } from '../lora.js';
|
|
16
|
-
import type { AttentionConfig, AttentionState, AttentionResult, AttentionDebugFlags } from './types.js';
|
|
17
|
-
|
|
18
|
-
/**
|
|
19
|
-
* Record attention for a single layer (batched, no submit).
|
|
20
|
-
*
|
|
21
|
-
* Uses record* kernel variants to batch all GPU operations into a shared
|
|
22
|
-
* command encoder. No GPU submits happen here - submit once at end of forward pass.
|
|
23
|
-
*/
|
|
24
|
-
export function recordLayerAttentionGPU(
|
|
25
|
-
recorder: CommandRecorder,
|
|
26
|
-
input: Tensor,
|
|
27
|
-
layerWeights: LayerWeights | null,
|
|
28
|
-
config: AttentionConfig,
|
|
29
|
-
state: AttentionState,
|
|
30
|
-
debug?: boolean,
|
|
31
|
-
debugFlags?: AttentionDebugFlags,
|
|
32
|
-
getWeightBuffer?: (weight: GPUBuffer | WeightBuffer | Float32Array | ArrayBuffer | CpuWeightBuffer, label: string) => GPUBuffer | WeightBuffer,
|
|
33
|
-
getNormWeightBuffer?: (weight: GPUBuffer | Float32Array | ArrayBuffer | CpuWeightBuffer, label: string) => GPUBuffer,
|
|
34
|
-
debugCheckBuffer?: (buffer: GPUBuffer, label: string, numTokens: number, expectedDim?: number) => Promise<void>,
|
|
35
|
-
lora?: LoRAAdapter | null
|
|
36
|
-
): Promise<AttentionResult>;
|
|
@@ -1,8 +0,0 @@
|
|
|
1
|
-
import type { LayerWeights } from '../types.js';
|
|
2
|
-
|
|
3
|
-
export declare function hasQ4KMaterialization(weight: unknown): boolean;
|
|
4
|
-
|
|
5
|
-
export declare function canUseRmsNormWideTileProjectionFusion(
|
|
6
|
-
layerWeights: LayerWeights | null | undefined,
|
|
7
|
-
reusesSharedKV: boolean
|
|
8
|
-
): boolean;
|
|
@@ -1,38 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* Attention Run - Immediate GPU submission path
|
|
3
|
-
*
|
|
4
|
-
* Contains runLayerAttentionGPU which executes attention operations
|
|
5
|
-
* with immediate GPU submission (each kernel submits independently).
|
|
6
|
-
*
|
|
7
|
-
* @module inference/pipelines/text/attention/run
|
|
8
|
-
*/
|
|
9
|
-
|
|
10
|
-
import type { LayerWeights } from '../types.js';
|
|
11
|
-
import type { WeightBuffer, CpuWeightBuffer } from '../../../../gpu/weight-buffer.js';
|
|
12
|
-
import type { Tensor } from '../../../../gpu/tensor.js';
|
|
13
|
-
import type { LoRAAdapter } from '../lora.js';
|
|
14
|
-
import type { AttentionConfig, AttentionState, AttentionResult, AttentionDebugFlags } from './types.js';
|
|
15
|
-
|
|
16
|
-
/**
|
|
17
|
-
* Run attention for a single layer (GPU path).
|
|
18
|
-
*
|
|
19
|
-
* @param input - Input hidden states tensor
|
|
20
|
-
* @param layerWeights - Weights for this layer
|
|
21
|
-
* @param config - Attention configuration
|
|
22
|
-
* @param state - Shared state (RoPE freqs, KV cache)
|
|
23
|
-
* @param debug - Debug mode flag
|
|
24
|
-
* @param debugFlags - Mutable debug flags to prevent repeated logging
|
|
25
|
-
* @returns Output tensor after attention
|
|
26
|
-
*/
|
|
27
|
-
export function runLayerAttentionGPU(
|
|
28
|
-
input: Tensor,
|
|
29
|
-
layerWeights: LayerWeights | null,
|
|
30
|
-
config: AttentionConfig,
|
|
31
|
-
state: AttentionState,
|
|
32
|
-
debug?: boolean,
|
|
33
|
-
debugFlags?: AttentionDebugFlags,
|
|
34
|
-
getWeightBuffer?: (weight: GPUBuffer | WeightBuffer | Float32Array | ArrayBuffer | CpuWeightBuffer, label: string) => GPUBuffer | WeightBuffer,
|
|
35
|
-
getNormWeightBuffer?: (weight: GPUBuffer | Float32Array | ArrayBuffer | CpuWeightBuffer, label: string) => GPUBuffer,
|
|
36
|
-
debugCheckBuffer?: (buffer: GPUBuffer, label: string, numTokens: number, expectedDim?: number) => Promise<void>,
|
|
37
|
-
lora?: LoRAAdapter | null
|
|
38
|
-
): Promise<AttentionResult>;
|
|
@@ -1,125 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* Attention Types and Utilities
|
|
3
|
-
*
|
|
4
|
-
* Shared interfaces, debug helpers, and utility functions for attention operations.
|
|
5
|
-
*
|
|
6
|
-
* @module inference/pipelines/text/attention/types
|
|
7
|
-
*/
|
|
8
|
-
|
|
9
|
-
import type { CommandRecorder } from '../../../../gpu/kernel-selector.js';
|
|
10
|
-
import type { KVCacheInterface } from '../types.js';
|
|
11
|
-
import type { Tensor } from '../../../../gpu/tensor.js';
|
|
12
|
-
import type { LinearAttentionRuntime } from '../linear-attention.js';
|
|
13
|
-
import type { ExecutionV1PoliciesSchema } from '../../../../config/schema/execution-v1.schema.js';
|
|
14
|
-
|
|
15
|
-
/**
|
|
16
|
-
* Attention configuration for a layer.
|
|
17
|
-
*/
|
|
18
|
-
export interface AttentionConfig {
|
|
19
|
-
layerIdx: number;
|
|
20
|
-
numTokens: number;
|
|
21
|
-
isPrefill: boolean;
|
|
22
|
-
numHeads: number;
|
|
23
|
-
numKVHeads: number;
|
|
24
|
-
headDim: number;
|
|
25
|
-
hiddenSize: number;
|
|
26
|
-
rmsNormEps: number;
|
|
27
|
-
currentSeqLen: number;
|
|
28
|
-
/** Expected activation dtype from runtime config. */
|
|
29
|
-
activationDtype?: 'f16' | 'f32';
|
|
30
|
-
/** Explicit KV-cache target dtype for this attention step. */
|
|
31
|
-
kvDtype?: 'f16' | 'f32';
|
|
32
|
-
slidingWindow?: number | null;
|
|
33
|
-
layerType?: string;
|
|
34
|
-
/** Residual tensor for fused o_proj + residual add (decode only) */
|
|
35
|
-
residualTensor?: Tensor | null;
|
|
36
|
-
/** Skip input RMSNorm even if weights are present */
|
|
37
|
-
skipInputNorm?: boolean;
|
|
38
|
-
/** Gemma 2 attention softcapping: score = tanh(score / softcap) * softcap. 0 = disabled. */
|
|
39
|
-
attnSoftcap?: number;
|
|
40
|
-
/** Gemma 2 attention scaling: uses head_dim (256) instead of sqrt(head_dim) (16). */
|
|
41
|
-
queryPreAttnScalar?: number;
|
|
42
|
-
/** Apply query/key RMSNorm where per-head weights are declared. */
|
|
43
|
-
queryKeyNorm?: boolean;
|
|
44
|
-
/** Layers that apply query/key RMSNorm; null means all layers when queryKeyNorm=true. */
|
|
45
|
-
queryKeyNormLayers?: number[] | null;
|
|
46
|
-
/** Layers that carry explicit Q/K RMSNorm scale tensors; null means missing weights are invalid. */
|
|
47
|
-
queryKeyNormWeightLayers?: number[] | null;
|
|
48
|
-
/** Apply unit-scale value RMSNorm before attention. */
|
|
49
|
-
valueNorm?: boolean;
|
|
50
|
-
/** Apply sigmoid gate from q_proj split to attention output before o_proj. */
|
|
51
|
-
attentionOutputGate?: boolean;
|
|
52
|
-
/** Gemma 2 RMS scaling: (1+w)*x */
|
|
53
|
-
rmsNormWeightOffset?: boolean;
|
|
54
|
-
/** Whether causal masking is enabled (default true). */
|
|
55
|
-
causalAttention?: boolean;
|
|
56
|
-
/** Absolute-position multimodal span that should remain bidirectional during causal prefill. */
|
|
57
|
-
multimodalBidirectionalSpan?: {
|
|
58
|
-
start: number;
|
|
59
|
-
length: number;
|
|
60
|
-
} | null;
|
|
61
|
-
/** RoPE rotary dimension (may differ from headDim with partial rotary). */
|
|
62
|
-
ropeRotaryDim?: number;
|
|
63
|
-
/** RoPE frequency base dimension; non-interleaved kernels use this as the rotate-half pair span. */
|
|
64
|
-
ropeFrequencyBaseDim?: number;
|
|
65
|
-
/** Whether RoPE uses interleaved layout. */
|
|
66
|
-
ropeInterleaved?: boolean;
|
|
67
|
-
/** Token IDs for the current micro-batch (required by BDPA KV ingestion). */
|
|
68
|
-
tokenIds?: number[] | null;
|
|
69
|
-
/** Kernel path override for attention dispatch. */
|
|
70
|
-
kernelPath?: Record<string, unknown> | null;
|
|
71
|
-
/** Disable RoPE for this layer (e.g., non-rotary attention). */
|
|
72
|
-
disableRoPE?: boolean;
|
|
73
|
-
}
|
|
74
|
-
|
|
75
|
-
/**
|
|
76
|
-
* Attention state passed between operations.
|
|
77
|
-
*/
|
|
78
|
-
export interface AttentionState {
|
|
79
|
-
ropeFreqsCos: GPUBuffer | null;
|
|
80
|
-
ropeFreqsSin: GPUBuffer | null;
|
|
81
|
-
kvCache: KVCacheInterface;
|
|
82
|
-
linearRuntime?: LinearAttentionRuntime | null;
|
|
83
|
-
executionPolicies?: ExecutionV1PoliciesSchema | null;
|
|
84
|
-
}
|
|
85
|
-
|
|
86
|
-
/**
|
|
87
|
-
* Result from attention layer execution.
|
|
88
|
-
*/
|
|
89
|
-
export interface AttentionResult {
|
|
90
|
-
/** Output tensor after attention + o_proj */
|
|
91
|
-
output: Tensor;
|
|
92
|
-
/** Whether the attention residual was fused into o_proj (layer.ts should skip residual add) */
|
|
93
|
-
residualFused: boolean;
|
|
94
|
-
}
|
|
95
|
-
|
|
96
|
-
/**
|
|
97
|
-
* Debug flags for attention - tracks which layer/stage combos have been logged.
|
|
98
|
-
* Uses a Set of "L{layer}_{stage}" keys to prevent duplicate logging.
|
|
99
|
-
*/
|
|
100
|
-
export interface AttentionDebugFlags {
|
|
101
|
-
/** Layers to debug (null = none, empty = layer 0 only for backward compat) */
|
|
102
|
-
debugLayers?: number[] | null;
|
|
103
|
-
/** Set of "L{layer}_{stage}" keys that have been logged */
|
|
104
|
-
loggedStages?: Set<string>;
|
|
105
|
-
}
|
|
106
|
-
|
|
107
|
-
/**
|
|
108
|
-
* Check if a layer should be debugged
|
|
109
|
-
*/
|
|
110
|
-
export function shouldDebugLayer(layerIdx: number, debugLayers: number[] | null | undefined): boolean;
|
|
111
|
-
|
|
112
|
-
/**
|
|
113
|
-
* Check if a stage has been logged for a layer, and mark it as logged
|
|
114
|
-
*/
|
|
115
|
-
export function markStageLogged(layerIdx: number, stage: string, flags: AttentionDebugFlags): boolean;
|
|
116
|
-
|
|
117
|
-
/**
|
|
118
|
-
* Release buffer or track for later cleanup (recording mode).
|
|
119
|
-
*/
|
|
120
|
-
export function releaseOrTrack(recorder: CommandRecorder | undefined, buffer: GPUBuffer): void;
|
|
121
|
-
|
|
122
|
-
/**
|
|
123
|
-
* Get or create a buffer of ones for Q/K norm when per-head weights are absent.
|
|
124
|
-
*/
|
|
125
|
-
export function getQKNormOnesBuffer(headDim: number): GPUBuffer;
|
|
@@ -1,23 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* Attention Module - Public aggregation entrypoint
|
|
3
|
-
*
|
|
4
|
-
* This file re-exports the attention public API from the attention/ directory.
|
|
5
|
-
*
|
|
6
|
-
* @module inference/pipelines/text/attention
|
|
7
|
-
*/
|
|
8
|
-
|
|
9
|
-
export {
|
|
10
|
-
// Types
|
|
11
|
-
type AttentionConfig,
|
|
12
|
-
type AttentionState,
|
|
13
|
-
type AttentionResult,
|
|
14
|
-
type AttentionDebugFlags,
|
|
15
|
-
// Utilities
|
|
16
|
-
shouldDebugLayer,
|
|
17
|
-
markStageLogged,
|
|
18
|
-
releaseOrTrack,
|
|
19
|
-
getQKNormOnesBuffer,
|
|
20
|
-
// Functions
|
|
21
|
-
runLayerAttentionGPU,
|
|
22
|
-
recordLayerAttentionGPU,
|
|
23
|
-
} from './attention/index.js';
|
|
@@ -1,22 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* BDPA steamroller host-side preprocessing helpers.
|
|
3
|
-
*/
|
|
4
|
-
|
|
5
|
-
export interface BDPAGeneratedData {
|
|
6
|
-
tBasisK: Float32Array;
|
|
7
|
-
tBasisV: Float32Array;
|
|
8
|
-
pDeltaK: Int8Array;
|
|
9
|
-
pDeltaV: Int8Array;
|
|
10
|
-
iFlat: Uint32Array;
|
|
11
|
-
numBasisVectors: number;
|
|
12
|
-
}
|
|
13
|
-
|
|
14
|
-
export declare function steamrollerRadixArgsort(tokens: Int32Array | Uint32Array): Uint32Array;
|
|
15
|
-
|
|
16
|
-
export declare function generateBDPAData(
|
|
17
|
-
sortedIndices: Uint32Array,
|
|
18
|
-
kBuffer: Float32Array | Uint16Array,
|
|
19
|
-
vBuffer: Float32Array | Uint16Array,
|
|
20
|
-
tokens: Int32Array | Uint32Array,
|
|
21
|
-
kvSize: number
|
|
22
|
-
): BDPAGeneratedData;
|