doppler-gpu 0.4.7 → 0.4.9
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 +74 -0
- package/README.md +93 -144
- package/models/catalog.json +1589 -0
- package/package.json +44 -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-3-5-9b-f16-af32.json +474 -0
- package/src/config/conversion/qwen3/qwen-3-5-9b-q4k-ehaf16.json +509 -0
- 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/backward-registry.json +13 -2
- package/src/config/kernels/kernel-ref-digests.js +34 -21
- package/src/config/kernels/registry.json +1402 -2623
- 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-3-5-9b-metal-correctness.json +46 -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/schema/training-metrics.schema.js +4 -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/bridge/trainer-artifact-bridge.d.ts +161 -0
- package/src/experimental/bridge/trainer-artifact-bridge.js +573 -0
- package/src/experimental/orchestration/index.d.ts +10 -1
- package/src/experimental/training/attention-backward.js +123 -55
- package/src/experimental/training/autograd.js +101 -5
- package/src/experimental/training/clip.js +6 -1
- package/src/experimental/training/datasets/text-pairs.js +80 -8
- package/src/experimental/training/distillation/student-fixture.js +470 -114
- package/src/experimental/training/lora-pipeline.js +202 -27
- package/src/experimental/training/lora.js +10 -4
- package/src/experimental/training/optimizer.js +9 -1
- package/src/experimental/training/runner.js +18 -4
- package/src/experimental/training/suite.js +1 -1
- package/src/experimental/training/workloads.js +11 -0
- 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/backward/adam.js +6 -4
- package/src/gpu/kernels/backward/causal_conv1d_silu_backward.d.ts +24 -0
- package/src/gpu/kernels/backward/causal_conv1d_silu_backward.js +74 -0
- package/src/gpu/kernels/backward/causal_conv1d_silu_backward.wgsl +59 -0
- package/src/gpu/kernels/backward/cross_entropy_backward.wgsl +4 -0
- package/src/gpu/kernels/backward/gated_delta_recurrent_backward.d.ts +39 -0
- package/src/gpu/kernels/backward/gated_delta_recurrent_backward.js +124 -0
- package/src/gpu/kernels/backward/gated_delta_recurrent_backward.wgsl +177 -0
- package/src/gpu/kernels/backward/gated_rmsnorm_backward.d.ts +32 -0
- package/src/gpu/kernels/backward/gated_rmsnorm_backward.js +90 -0
- package/src/gpu/kernels/backward/gated_rmsnorm_backward.wgsl +86 -0
- package/src/gpu/kernels/backward/index.d.ts +3 -0
- package/src/gpu/kernels/backward/index.js +3 -0
- package/src/gpu/kernels/backward/matmul_backward.wgsl +6 -6
- package/src/gpu/kernels/backward/matmul_backward_f16w.wgsl +82 -0
- package/src/gpu/kernels/backward/rmsnorm_backward.js +26 -4
- package/src/gpu/kernels/backward/rmsnorm_backward.wgsl +6 -3
- package/src/gpu/kernels/backward/utils.d.ts +12 -2
- package/src/gpu/kernels/backward/utils.js +9 -2
- package/src/gpu/kernels/cast_f16_to_f32.wgsl +6 -2
- 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 +26 -0
- package/src/gpu/kernels/index.js +18 -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 +10 -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.d.ts +1 -0
- package/src/gpu/kernels/silu.js +19 -4
- 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/embed.js +1 -1
- 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 +130 -10
- package/src/inference/pipelines/text/ffn/dense.js +241 -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 +326 -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/lora-apply.js +4 -4
- package/src/inference/pipelines/text/model-load.js +6 -0
- package/src/inference/pipelines/text/moe-cpu.js +1 -0
- package/src/inference/pipelines/text/moe-gpu.js +1 -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 +150 -55
- 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 +2 -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/node-converter.js +20 -5
- package/src/tooling/opfs-cache.d.ts +21 -1
- package/src/tooling/opfs-cache.js +168 -32
- package/src/tooling-experimental-exports.shared.d.ts +25 -0
- package/src/tooling-experimental-exports.shared.js +18 -0
- 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
|
@@ -1,49 +0,0 @@
|
|
|
1
|
-
export async function withMappedReadBuffer(buffer, read) {
|
|
2
|
-
let mapped = false;
|
|
3
|
-
try {
|
|
4
|
-
await buffer.mapAsync(GPUMapMode.READ);
|
|
5
|
-
mapped = true;
|
|
6
|
-
return await read(buffer.getMappedRange());
|
|
7
|
-
} finally {
|
|
8
|
-
if (mapped) {
|
|
9
|
-
buffer.unmap();
|
|
10
|
-
}
|
|
11
|
-
}
|
|
12
|
-
}
|
|
13
|
-
|
|
14
|
-
export async function withMappedReadBuffers(entries, read) {
|
|
15
|
-
const mappedEntries = [];
|
|
16
|
-
let mapError = null;
|
|
17
|
-
try {
|
|
18
|
-
await Promise.all(entries.map(async (entry) => {
|
|
19
|
-
try {
|
|
20
|
-
await entry.buffer.mapAsync(GPUMapMode.READ);
|
|
21
|
-
mappedEntries.push(entry);
|
|
22
|
-
} catch (error) {
|
|
23
|
-
mapError ??= error;
|
|
24
|
-
}
|
|
25
|
-
}));
|
|
26
|
-
if (mapError) {
|
|
27
|
-
throw mapError;
|
|
28
|
-
}
|
|
29
|
-
const ranges = entries.map((entry) => {
|
|
30
|
-
if (entry.offset != null || entry.size != null) {
|
|
31
|
-
return entry.buffer.getMappedRange(
|
|
32
|
-
entry.offset ?? 0,
|
|
33
|
-
entry.size ?? (entry.buffer.size - (entry.offset ?? 0))
|
|
34
|
-
);
|
|
35
|
-
}
|
|
36
|
-
return entry.buffer.getMappedRange();
|
|
37
|
-
});
|
|
38
|
-
return await read(ranges);
|
|
39
|
-
} finally {
|
|
40
|
-
for (let index = mappedEntries.length - 1; index >= 0; index -= 1) {
|
|
41
|
-
mappedEntries[index].buffer.unmap();
|
|
42
|
-
}
|
|
43
|
-
for (const entry of entries) {
|
|
44
|
-
if (entry.destroy === true) {
|
|
45
|
-
entry.buffer.destroy();
|
|
46
|
-
}
|
|
47
|
-
}
|
|
48
|
-
}
|
|
49
|
-
}
|
|
@@ -1,12 +0,0 @@
|
|
|
1
|
-
/**
|
|
2
|
-
* GPU submit latency probe.
|
|
3
|
-
*
|
|
4
|
-
* Dispatches a trivial compute shader, submits, waits for the GPU fence
|
|
5
|
-
* plus a staging-buffer readback, and returns the wall-clock roundtrip in ms.
|
|
6
|
-
*/
|
|
7
|
-
|
|
8
|
-
/**
|
|
9
|
-
* Run a single submit+readback roundtrip and return the elapsed time in ms.
|
|
10
|
-
* Returns `null` if the probe cannot run.
|
|
11
|
-
*/
|
|
12
|
-
export function probeSubmitLatency(device: GPUDevice): Promise<number | null>;
|
package/src/index-browser.d.ts
DELETED
package/src/index-browser.js
DELETED
package/src/inference/README.md
DELETED
|
@@ -1,593 +0,0 @@
|
|
|
1
|
-
# Inference Module
|
|
2
|
-
|
|
3
|
-
Purpose: End-to-end inference pipeline orchestration and helpers.
|
|
4
|
-
|
|
5
|
-
Boundary note:
|
|
6
|
-
- Contract boundaries are canonical in `docs/pipeline-contract.md`.
|
|
7
|
-
- This file is implementation-level reference for inference internals.
|
|
8
|
-
|
|
9
|
-
## Scope
|
|
10
|
-
|
|
11
|
-
- Pipeline orchestration, tokenization, KV cache, MoE routing, sampling.
|
|
12
|
-
- Modular pipeline helpers and refactor checklist context.
|
|
13
|
-
|
|
14
|
-
This directory contains DOPPLER's inference pipeline and its main building blocks:
|
|
15
|
-
|
|
16
|
-
- `pipelines/text.js`: end-to-end text orchestration (prefill, decode loop, sampling, debug)
|
|
17
|
-
- `pipelines/diffusion/`: image diffusion pipeline modules
|
|
18
|
-
- `pipelines/energy/`: energy-model pipeline modules
|
|
19
|
-
- `tokenizer.js`: tokenizer backends and manifest-driven selection
|
|
20
|
-
- `kv-cache.js`: KV cache layouts and policies (full and sliding-window)
|
|
21
|
-
- `moe-router.js`: MoE routing and expert execution planning
|
|
22
|
-
- `speculative.js`: speculative decoding scaffolding
|
|
23
|
-
- `pipelines/text/`: modular text helpers (split for maintainability)
|
|
24
|
-
- **WIRED**: `config.js`, `init.js`, `generator.js`, `generator-steps.js`,
|
|
25
|
-
`embed.js`, `layer.js`, `logits.js`, `sampling.js`, `attention/`, `ffn/`
|
|
26
|
-
- Modular entrypoints are used by `pipelines/text.js` + `pipelines/text/generator.js`
|
|
27
|
-
- No separate prefill/decode modules are required in current layout
|
|
28
|
-
|
|
29
|
-
## Related
|
|
30
|
-
- [Line-by-Line Pipeline Steps](#line-by-line-pipeline-steps) below for detailed execution map
|
|
31
|
-
- `docs/architecture.md` for system terminology and runtime dataflow context
|
|
32
|
-
- `gpu/kernel-selector.js` for the WGSL wrapper API (`run*` and `record*`)
|
|
33
|
-
|
|
34
|
-
---
|
|
35
|
-
|
|
36
|
-
## The Full Inference Graph (Conceptual)
|
|
37
|
-
|
|
38
|
-
This is the high-level execution graph for text generation. It is a dataflow graph with state (KV cache) plus a decode loop.
|
|
39
|
-
|
|
40
|
-
```dot
|
|
41
|
-
digraph DopplerInference {
|
|
42
|
-
rankdir=LR;
|
|
43
|
-
compound=true;
|
|
44
|
-
node [shape=box, fontsize=10];
|
|
45
|
-
|
|
46
|
-
subgraph cluster_cpu {
|
|
47
|
-
label="CPU (JS orchestration)";
|
|
48
|
-
style=rounded;
|
|
49
|
-
|
|
50
|
-
prompt [label="prompt (string)"];
|
|
51
|
-
tokenize [label="Tokenizer.encode()"];
|
|
52
|
-
opts [label="generation options\n(temperature/topK/topP/stop)"];
|
|
53
|
-
sample [label="sampling decision\n(topK/topP/temperature)\n(or greedy)"];
|
|
54
|
-
detok [label="Tokenizer.decode(token)"];
|
|
55
|
-
stop [label="stop checks\n(stopTokens/stopSequences/maxTokens)"];
|
|
56
|
-
stream [label="emit token text"];
|
|
57
|
-
}
|
|
58
|
-
|
|
59
|
-
subgraph cluster_storage {
|
|
60
|
-
label="Storage (weights + assets)";
|
|
61
|
-
style=rounded;
|
|
62
|
-
|
|
63
|
-
manifest [label="manifest.json\n(RDRR)"];
|
|
64
|
-
shards [label="shard_*.bin\n(RDRR)"];
|
|
65
|
-
opfs [label="OPFS cache"];
|
|
66
|
-
bridge [label="Native Bridge\n(optional)"];
|
|
67
|
-
http [label="HTTP fetch\n(optional)"];
|
|
68
|
-
}
|
|
69
|
-
|
|
70
|
-
subgraph cluster_gpu {
|
|
71
|
-
label="GPU (WebGPU / WGSL kernels)";
|
|
72
|
-
style=rounded;
|
|
73
|
-
|
|
74
|
-
embed [label="runGather\n(embedding)"];
|
|
75
|
-
layerloop [label="for each layer l:\nattention + ffn"];
|
|
76
|
-
lmhead [label="runMatmul\n(LM head -> logits)"];
|
|
77
|
-
kv [label="KV cache (GPU buffers)\nstateful"];
|
|
78
|
-
}
|
|
79
|
-
|
|
80
|
-
prompt -> tokenize;
|
|
81
|
-
opts -> tokenize;
|
|
82
|
-
|
|
83
|
-
manifest -> tokenize [style=dashed, label="tokenizer config"];
|
|
84
|
-
manifest -> layerloop [style=dashed, label="architecture/config"];
|
|
85
|
-
manifest -> shards [style=dashed];
|
|
86
|
-
|
|
87
|
-
shards -> opfs [label="store"];
|
|
88
|
-
http -> shards [label="download"];
|
|
89
|
-
bridge -> shards [label="read local"];
|
|
90
|
-
|
|
91
|
-
tokenize -> embed [label="tokenIds"];
|
|
92
|
-
embed -> layerloop;
|
|
93
|
-
|
|
94
|
-
layerloop -> kv [label="write K/V"];
|
|
95
|
-
kv -> layerloop [label="read K/V"];
|
|
96
|
-
|
|
97
|
-
layerloop -> lmhead;
|
|
98
|
-
lmhead -> sample [label="logits readback\n(sync point)"];
|
|
99
|
-
|
|
100
|
-
sample -> detok;
|
|
101
|
-
detok -> stream;
|
|
102
|
-
stream -> stop;
|
|
103
|
-
stop -> layerloop [label="decode loop\n(next token)"];
|
|
104
|
-
}
|
|
105
|
-
```
|
|
106
|
-
|
|
107
|
-
CPU and GPU boundary notes:
|
|
108
|
-
|
|
109
|
-
- `device.queue.submit()` schedules GPU work and returns immediately.
|
|
110
|
-
- `readBuffer(...)` for logits (or any GPU readback) creates a synchronization point.
|
|
111
|
-
- The KV cache is GPU-resident state that persists across decode steps.
|
|
112
|
-
|
|
113
|
-
---
|
|
114
|
-
|
|
115
|
-
## One Transformer Layer Graph (Kernel Names)
|
|
116
|
-
|
|
117
|
-
The per-layer graph below uses the kernel wrapper names from `gpu/kernel-selector.js`. The actual WGSL variant selected depends on device features (`shader-f16`, `subgroups`) and runtime choices (attention tier, batching).
|
|
118
|
-
|
|
119
|
-
```dot
|
|
120
|
-
digraph DopplerLayer {
|
|
121
|
-
rankdir=LR;
|
|
122
|
-
node [shape=box, fontsize=10];
|
|
123
|
-
|
|
124
|
-
x_in [label="hidden states X\n(GPUBuffer)"];
|
|
125
|
-
|
|
126
|
-
norm1 [label="runRMSNorm"];
|
|
127
|
-
q [label="runMatmul (Q)"];
|
|
128
|
-
k [label="runMatmul (K)"];
|
|
129
|
-
v [label="runMatmul (V)"];
|
|
130
|
-
qnorm [label="runRMSNorm (Q)\n(optional)"];
|
|
131
|
-
knorm [label="runRMSNorm (K)\n(optional)"];
|
|
132
|
-
rope_q [label="runRoPE (Q)"];
|
|
133
|
-
rope_k [label="runRoPE (K)"];
|
|
134
|
-
kv_write [label="KV cache update\n(+ castF32ToF16 if enabled)"];
|
|
135
|
-
attn [label="runAttention\n(tiled_large/tiled_small/streaming,\n_prefill or _decode variant)"];
|
|
136
|
-
o [label="runMatmul (O)"];
|
|
137
|
-
res1 [label="runResidualAdd"];
|
|
138
|
-
|
|
139
|
-
norm2 [label="runRMSNorm\n(model-dependent)"];
|
|
140
|
-
|
|
141
|
-
ffn_dense [label="Dense FFN path:\nrunMatmul -> runSiLU/runGeLU\n(or runSwiGLURowsplitBias)\n-> runMatmul"];
|
|
142
|
-
ffn_moe [label="MoE FFN path:\ncomputeRouterLogitsGPU (uses runMatmul)\n-> runSoftmaxTopK\n-> runMoEGather\n-> expert matmuls (+ dequantizeMXFP4Expert)\n-> runScatterAddDynamic"];
|
|
143
|
-
|
|
144
|
-
res2 [label="runResidualAdd"];
|
|
145
|
-
x_out [label="hidden states X'\n(GPUBuffer)"];
|
|
146
|
-
|
|
147
|
-
x_in -> norm1 -> q;
|
|
148
|
-
norm1 -> k;
|
|
149
|
-
norm1 -> v;
|
|
150
|
-
|
|
151
|
-
q -> qnorm [style=dashed, label="if qkNorm"];
|
|
152
|
-
k -> knorm [style=dashed, label="if qkNorm"];
|
|
153
|
-
qnorm -> rope_q;
|
|
154
|
-
knorm -> rope_k;
|
|
155
|
-
q -> rope_q [style=dashed, label="else"];
|
|
156
|
-
k -> rope_k [style=dashed, label="else"];
|
|
157
|
-
|
|
158
|
-
rope_k -> kv_write;
|
|
159
|
-
v -> kv_write;
|
|
160
|
-
|
|
161
|
-
rope_q -> attn;
|
|
162
|
-
kv_write -> attn;
|
|
163
|
-
|
|
164
|
-
attn -> o -> res1;
|
|
165
|
-
x_in -> res1 [style=dashed, label="skip"];
|
|
166
|
-
|
|
167
|
-
res1 -> norm2;
|
|
168
|
-
norm2 -> ffn_dense;
|
|
169
|
-
norm2 -> ffn_moe;
|
|
170
|
-
|
|
171
|
-
ffn_dense -> res2;
|
|
172
|
-
ffn_moe -> res2;
|
|
173
|
-
res1 -> res2 [style=dashed, label="skip"];
|
|
174
|
-
|
|
175
|
-
res2 -> x_out;
|
|
176
|
-
}
|
|
177
|
-
```
|
|
178
|
-
|
|
179
|
-
Notes:
|
|
180
|
-
|
|
181
|
-
- Prefill and decode share the same graph, but use different attention behavior:
|
|
182
|
-
- Prefill writes many KV entries and can use a different attention kernel path.
|
|
183
|
-
- Decode appends one KV entry per step and reads the full cache for attention.
|
|
184
|
-
- Many kernels have both `run*` and `record*` forms:
|
|
185
|
-
- `run*` submits immediately
|
|
186
|
-
- `record*` records into a shared command buffer for batching
|
|
187
|
-
|
|
188
|
-
---
|
|
189
|
-
|
|
190
|
-
## Where WGSL Lives
|
|
191
|
-
|
|
192
|
-
WGSL kernels are in `gpu/kernels/`. The wrappers in `gpu/kernel-selector.js` compile and dispatch them, selecting variants based on device capabilities and the operation's dtype.
|
|
193
|
-
|
|
194
|
-
---
|
|
195
|
-
|
|
196
|
-
## Line-by-Line Pipeline Steps
|
|
197
|
-
|
|
198
|
-
### Legend
|
|
199
|
-
```
|
|
200
|
-
[STEP] = Required step
|
|
201
|
-
[STEP]? = Conditional step
|
|
202
|
-
├─ FORK = Model-specific branch
|
|
203
|
-
├─ DEP = External dependency
|
|
204
|
-
└─ SKIP = Step skipped under condition
|
|
205
|
-
```
|
|
206
|
-
|
|
207
|
-
---
|
|
208
|
-
|
|
209
|
-
### PHASE 0: INITIALIZATION
|
|
210
|
-
|
|
211
|
-
```
|
|
212
|
-
[0.1] getMemoryCapabilities()
|
|
213
|
-
├─ DEP: Browser Memory64 support (Chrome 111+, Firefox 119+)
|
|
214
|
-
├─ DEP: Unified memory detection (Apple Silicon, AMD Strix)
|
|
215
|
-
└─ Output: { hasMemory64, isUnifiedMemory, maxHeapSize }
|
|
216
|
-
|
|
217
|
-
[0.2] initDevice()
|
|
218
|
-
├─ DEP: WebGPU adapter availability
|
|
219
|
-
├─ DEP: GPU features: shader-f16, subgroups, timestamp-query
|
|
220
|
-
└─ Output: { device, hasF16, hasSubgroups, limits }
|
|
221
|
-
|
|
222
|
-
[0.3] loadManifestFromStore()
|
|
223
|
-
├─ DEP: OPFS support (Chrome 102+)
|
|
224
|
-
├─ ALT: Native Bridge (if storageContext.loadShard provided)
|
|
225
|
-
└─ Output: manifest JSON
|
|
226
|
-
```
|
|
227
|
-
|
|
228
|
-
---
|
|
229
|
-
|
|
230
|
-
### PHASE 1: MODEL LOADING
|
|
231
|
-
|
|
232
|
-
```
|
|
233
|
-
[1.1] DopplerLoader.init()
|
|
234
|
-
├─ FORK: Dense model on discrete GPU → Warning (10x slowdown expected)
|
|
235
|
-
├─ FORK: MoE model → Set useMoE flag
|
|
236
|
-
└─ Output: Loader ready
|
|
237
|
-
|
|
238
|
-
[1.2] _buildTensorLocations()
|
|
239
|
-
└─ Output: Map<tensorName, {shard, offset, size, shape, dtype}>
|
|
240
|
-
|
|
241
|
-
[1.3] _loadEmbeddings()
|
|
242
|
-
├─ FORK: Gemma 3 → "language_model.model.embed_tokens.weight"
|
|
243
|
-
├─ FORK: LLaMA/Mistral → "model.embed_tokens.weight"
|
|
244
|
-
├─ FORK: GPT → "wte.weight"
|
|
245
|
-
├─ DEP: GPU available → Upload to GPU buffer
|
|
246
|
-
└─ Output: embeddings (GPUBuffer or Float32Array)
|
|
247
|
-
|
|
248
|
-
[1.4] _loadLayer(layerIdx) × numLayers
|
|
249
|
-
│
|
|
250
|
-
├─[1.4.1] Load inputNorm
|
|
251
|
-
│ └─ "input_layernorm.weight"
|
|
252
|
-
│
|
|
253
|
-
├─[1.4.2] Load Q/K/V/O projections
|
|
254
|
-
│ ├─ FORK: Gemma 3 → "language_model.model.layers.{l}.self_attn.*"
|
|
255
|
-
│ └─ FORK: LLaMA → "model.layers.{l}.self_attn.*"
|
|
256
|
-
│
|
|
257
|
-
├─[1.4.3] Load Q/K norms?
|
|
258
|
-
│ ├─ FORK: Gemma → Load "self_attn.q_norm.weight", "self_attn.k_norm.weight"
|
|
259
|
-
│ └─ SKIP: LLaMA/Mistral (no Q/K norms)
|
|
260
|
-
│
|
|
261
|
-
├─[1.4.4] Load postAttentionNorm?
|
|
262
|
-
│ ├─ FORK: Gemma 3 → "post_attention_layernorm.weight"
|
|
263
|
-
│ └─ SKIP: LLaMA (uses postNorm = post_attention_layernorm before FFN)
|
|
264
|
-
│
|
|
265
|
-
├─[1.4.5] Load preFeedforwardNorm?
|
|
266
|
-
│ ├─ FORK: Gemma 3 → "pre_feedforward_layernorm.weight"
|
|
267
|
-
│ └─ SKIP: LLaMA/Mistral
|
|
268
|
-
│
|
|
269
|
-
├─[1.4.6] Load postFeedforwardNorm?
|
|
270
|
-
│ ├─ FORK: Gemma 3 → "post_feedforward_layernorm.weight"
|
|
271
|
-
│ └─ SKIP: LLaMA/Mistral
|
|
272
|
-
│
|
|
273
|
-
├─[1.4.7] Load FFN weights (Dense only)
|
|
274
|
-
│ ├─ FORK: Dense layer → Load gate/up/down
|
|
275
|
-
│ └─ SKIP: MoE layer (experts loaded on demand)
|
|
276
|
-
│
|
|
277
|
-
└─[1.4.8] F16 downcast?
|
|
278
|
-
├─ DEP: hasF16 = true → Downcast Q/K/V/O/gate/up/down to F16
|
|
279
|
-
└─ SKIP: No F16 support → Keep F32
|
|
280
|
-
|
|
281
|
-
[1.5] loadExpert(layerIdx, expertIdx)? [MoE only, on-demand]
|
|
282
|
-
├─ FORK: MoE model → Load when router selects expert
|
|
283
|
-
└─ SKIP: Dense model
|
|
284
|
-
|
|
285
|
-
[1.6] _loadFinalWeights()
|
|
286
|
-
├─[1.6.1] Load finalNorm
|
|
287
|
-
│ ├─ FORK: Gemma 3 → "language_model.model.norm.weight"
|
|
288
|
-
│ └─ FORK: LLaMA → "model.norm.weight"
|
|
289
|
-
│
|
|
290
|
-
└─[1.6.2] Load lmHead
|
|
291
|
-
├─ FORK: Tied embeddings → Reuse embeddings buffer (transposeB=true)
|
|
292
|
-
└─ FORK: Untied → Load "lm_head.weight"
|
|
293
|
-
|
|
294
|
-
[1.7] Dequantization (during tensor load)
|
|
295
|
-
├─ FORK: Q4_K_M → dequantize() kernel
|
|
296
|
-
│ ├─ DEP: hasSubgroups → dequant_subgroup (faster)
|
|
297
|
-
│ └─ ALT: dequant_shared (fallback)
|
|
298
|
-
├─ FORK: BF16 → bf16_to_f32 kernel
|
|
299
|
-
├─ FORK: F16 + hasF16 → Direct GPU upload
|
|
300
|
-
└─ FORK: F32 → Direct GPU upload
|
|
301
|
-
```
|
|
302
|
-
|
|
303
|
-
---
|
|
304
|
-
|
|
305
|
-
### PHASE 2: ROPE FREQUENCIES
|
|
306
|
-
|
|
307
|
-
```
|
|
308
|
-
[2.1] _initRoPEFrequencies()
|
|
309
|
-
├─ CONFIG: ropeTheta (default 10000, Gemma3=1000000)
|
|
310
|
-
├─ CONFIG: ropeScale (context extension factor)
|
|
311
|
-
├─ CONFIG: headDim (Gemma3=256, LLaMA=128)
|
|
312
|
-
└─ Output: ropeFreqsCos, ropeFreqsSin (GPUBuffer)
|
|
313
|
-
```
|
|
314
|
-
|
|
315
|
-
---
|
|
316
|
-
|
|
317
|
-
### PHASE 3: KV CACHE
|
|
318
|
-
|
|
319
|
-
```
|
|
320
|
-
[3.1] KVCache initialization
|
|
321
|
-
├─ FORK: slidingWindow > 0 → SlidingWindowKVCache
|
|
322
|
-
│ └─ CONFIG: Mistral slidingWindow=4096
|
|
323
|
-
└─ FORK: No window → Standard KVCache
|
|
324
|
-
|
|
325
|
-
├─ DEP: hasF16 → F16 KV cache (50% memory)
|
|
326
|
-
└─ ALT: F32 KV cache
|
|
327
|
-
|
|
328
|
-
└─ Output: Per-layer K,V buffers [maxSeqLen × numKVHeads × headDim]
|
|
329
|
-
```
|
|
330
|
-
|
|
331
|
-
---
|
|
332
|
-
|
|
333
|
-
### PHASE 4: TOKENIZER
|
|
334
|
-
|
|
335
|
-
```
|
|
336
|
-
[4.1] Tokenizer.init()
|
|
337
|
-
├─ Load tokenizer.json (bundled or network)
|
|
338
|
-
└─ Output: encode/decode functions
|
|
339
|
-
```
|
|
340
|
-
|
|
341
|
-
---
|
|
342
|
-
|
|
343
|
-
### PHASE 5: GENERATION START
|
|
344
|
-
|
|
345
|
-
```
|
|
346
|
-
[5.1] generate(prompt, options)
|
|
347
|
-
│
|
|
348
|
-
├─[5.1.1] Apply chat template?
|
|
349
|
-
│ ├─ FORK: modelConfig.chatTemplateType is set
|
|
350
|
-
│ │ → applyChatTemplate(prompt, type)
|
|
351
|
-
│ │ (gemma / llama3 / gpt-oss / chatml / qwen)
|
|
352
|
-
│ └─ SKIP: chatTemplateType unset (raw prompt)
|
|
353
|
-
│
|
|
354
|
-
└─[5.1.2] Tokenize
|
|
355
|
-
└─ Output: inputIds[]
|
|
356
|
-
```
|
|
357
|
-
|
|
358
|
-
---
|
|
359
|
-
|
|
360
|
-
### PHASE 6: PREFILL (Process entire prompt)
|
|
361
|
-
|
|
362
|
-
```
|
|
363
|
-
[6.1] _embed(inputIds)
|
|
364
|
-
│
|
|
365
|
-
├─[6.1.1] runGather(tokenIds, embeddings)
|
|
366
|
-
│ ├─ DEP: GPU → Zero-copy gather kernel
|
|
367
|
-
│ └─ ALT: CPU loop
|
|
368
|
-
│
|
|
369
|
-
└─[6.1.2] Scale embeddings?
|
|
370
|
-
├─ FORK: Gemma → embeddings *= sqrt(hiddenSize)
|
|
371
|
-
└─ SKIP: LLaMA/Mistral
|
|
372
|
-
|
|
373
|
-
[6.2] FOR l = 0 to numLayers-1:
|
|
374
|
-
│
|
|
375
|
-
└─ _processLayerGPU(l, hiddenStates, numTokens, isPrefill=true)
|
|
376
|
-
│
|
|
377
|
-
├─[6.2.1] Detect norm architecture
|
|
378
|
-
│ ├─ useSandwichNorm = (preFeedforwardNorm || postFeedforwardNorm) exists
|
|
379
|
-
│ ├─ FORK: Gemma 3 → useSandwichNorm = true
|
|
380
|
-
│ └─ FORK: LLaMA/Mistral → useSandwichNorm = false
|
|
381
|
-
│
|
|
382
|
-
├─[6.2.2] _attentionGPU()
|
|
383
|
-
│ │
|
|
384
|
-
│ ├─[6.2.2.1] runRMSNorm(input, inputNorm)
|
|
385
|
-
│ │
|
|
386
|
-
│ ├─[6.2.2.2] Q = runMatmul(normed, qProj, transposeB=true)
|
|
387
|
-
│ ├─[6.2.2.3] K = runMatmul(normed, kProj, transposeB=true)
|
|
388
|
-
│ ├─[6.2.2.4] V = runMatmul(normed, vProj, transposeB=true)
|
|
389
|
-
│ │
|
|
390
|
-
│ ├─[6.2.2.5] Q/K norms?
|
|
391
|
-
│ │ ├─ FORK: Gemma → runRMSNorm(Q, qNorm), runRMSNorm(K, kNorm)
|
|
392
|
-
│ │ └─ SKIP: LLaMA/Mistral
|
|
393
|
-
│ │
|
|
394
|
-
│ ├─[6.2.2.6] runRoPE(Q, K, freqsCos, freqsSin)
|
|
395
|
-
│ │
|
|
396
|
-
│ ├─[6.2.2.7] runAttention(Q, K, V, kvCache)
|
|
397
|
-
│ │ ├─ DEP: sharedMem >= 49KB → prefill_large (tileSize=64)
|
|
398
|
-
│ │ ├─ DEP: sharedMem >= 8KB → prefill_small (tileSize=32)
|
|
399
|
-
│ │ └─ ALT: prefill_streaming (no shared mem)
|
|
400
|
-
│ │ ├─ DEP: hasF16 → *_f16kv variant (F16 KV in shared mem)
|
|
401
|
-
│ │ └─ Output: attnOutput
|
|
402
|
-
│ │
|
|
403
|
-
│ └─[6.2.2.8] output = runMatmul(attnOutput, oProj, transposeB=true)
|
|
404
|
-
│
|
|
405
|
-
│ ┌─────────────────────────────────────────────────────────────┐
|
|
406
|
-
│ │ FORK: useSandwichNorm (Gemma 3) │
|
|
407
|
-
│ ├─────────────────────────────────────────────────────────────┤
|
|
408
|
-
├─[6.2.3a] postAttnNormed = runRMSNorm(attnOutput, postAttentionNorm)
|
|
409
|
-
├─[6.2.4a] postAttn = runResidualAdd(postAttnNormed, input) ← RESIDUAL AFTER NORM
|
|
410
|
-
├─[6.2.5a] ffnInput = runRMSNorm(postAttn, preFeedforwardNorm)
|
|
411
|
-
├─[6.2.6a] ffnOutput = _feedForwardGPU() or _moeFeedForwardGPU()
|
|
412
|
-
├─[6.2.7a] ffnNormed = runRMSNorm(ffnOutput, postFeedforwardNorm)
|
|
413
|
-
└─[6.2.8a] output = runResidualAdd(ffnNormed, postAttn) ← RESIDUAL AFTER NORM
|
|
414
|
-
│ └─────────────────────────────────────────────────────────────┘
|
|
415
|
-
│
|
|
416
|
-
│ ┌─────────────────────────────────────────────────────────────┐
|
|
417
|
-
│ │ FORK: Standard (LLaMA/Mistral) │
|
|
418
|
-
│ ├─────────────────────────────────────────────────────────────┤
|
|
419
|
-
├─[6.2.3b] postAttn = runResidualAdd(attnOutput, input) ← RESIDUAL FIRST
|
|
420
|
-
├─[6.2.4b] normed = runRMSNorm(postAttn, postAttnNorm)
|
|
421
|
-
├─[6.2.5b] ffnOutput = _feedForwardGPU() or _moeFeedForwardGPU()
|
|
422
|
-
└─[6.2.6b] output = runResidualAdd(ffnOutput, postAttn) ← RESIDUAL AFTER FFN
|
|
423
|
-
│ └─────────────────────────────────────────────────────────────┘
|
|
424
|
-
│
|
|
425
|
-
│ ┌─────────────────────────────────────────────────────────────┐
|
|
426
|
-
│ │ _feedForwardGPU (Dense) │
|
|
427
|
-
│ ├─────────────────────────────────────────────────────────────┤
|
|
428
|
-
│ ├─ gate = runMatmul(input, ffnGate, transposeB=true)
|
|
429
|
-
│ ├─ up = runMatmul(input, ffnUp, transposeB=true)
|
|
430
|
-
│ ├─ gated = runSiLU(gate) * up
|
|
431
|
-
│ └─ output = runMatmul(gated, ffnDown, transposeB=true)
|
|
432
|
-
│ └─────────────────────────────────────────────────────────────┘
|
|
433
|
-
│
|
|
434
|
-
│ ┌─────────────────────────────────────────────────────────────┐
|
|
435
|
-
│ │ _moeFeedForwardGPU (MoE) [Mixtral, etc.] │
|
|
436
|
-
│ ├─────────────────────────────────────────────────────────────┤
|
|
437
|
-
│ ├─ routerLogits = runMatmul(input, routerWeight)
|
|
438
|
-
│ ├─ (expertIndices, expertWeights) = runSoftmaxTopK(routerLogits, topK)
|
|
439
|
-
│ ├─ FOR each selected expert:
|
|
440
|
-
│ │ ├─ loadExpert(layerIdx, expertIdx) if not cached
|
|
441
|
-
│ │ ├─ gathered = runMoEGather(input, expertIndices)
|
|
442
|
-
│ │ ├─ expertOut = _feedForwardGPU(gathered, expertWeights)
|
|
443
|
-
│ │ └─ output = runScatterAdd(output, expertOut, expertIndices)
|
|
444
|
-
│ └─────────────────────────────────────────────────────────────┘
|
|
445
|
-
│
|
|
446
|
-
└─ Output: hiddenStates for next layer
|
|
447
|
-
|
|
448
|
-
[6.3] _computeLogits(finalHiddenStates, numTokens)
|
|
449
|
-
│
|
|
450
|
-
├─[6.3.1] runRMSNorm(hiddenStates, finalNorm)
|
|
451
|
-
│
|
|
452
|
-
├─[6.3.2] logits = runMatmul(normed, lmHead)
|
|
453
|
-
│ ├─ FORK: Tied embeddings → transposeB=true
|
|
454
|
-
│ └─ FORK: Untied → transposeB=true (SafeTensors layout)
|
|
455
|
-
│
|
|
456
|
-
└─[6.3.3] Extract last position logits
|
|
457
|
-
└─ Output: logits[vocabSize] for position (numTokens-1)
|
|
458
|
-
```
|
|
459
|
-
|
|
460
|
-
---
|
|
461
|
-
|
|
462
|
-
### PHASE 7: SAMPLING
|
|
463
|
-
|
|
464
|
-
```
|
|
465
|
-
[7.1] _applyRepetitionPenalty(logits, generatedIds, penalty)
|
|
466
|
-
└─ logits[token] /= penalty for each token in history
|
|
467
|
-
|
|
468
|
-
[7.2] _sample(logits, opts)
|
|
469
|
-
│
|
|
470
|
-
├─[7.2.1] Temperature scaling
|
|
471
|
-
│ └─ logits /= temperature
|
|
472
|
-
│
|
|
473
|
-
├─[7.2.2] Top-K?
|
|
474
|
-
│ ├─ CONFIG: topK > 0 → Apply
|
|
475
|
-
│ └─ SKIP: topK = 0
|
|
476
|
-
│
|
|
477
|
-
├─[7.2.3] Top-P (nucleus)?
|
|
478
|
-
│ ├─ CONFIG: topP < 1.0 → Apply
|
|
479
|
-
│ └─ SKIP: topP = 1.0
|
|
480
|
-
│
|
|
481
|
-
└─[7.2.4] Softmax + categorical sample
|
|
482
|
-
└─ Output: nextToken
|
|
483
|
-
```
|
|
484
|
-
|
|
485
|
-
---
|
|
486
|
-
|
|
487
|
-
### PHASE 8: DECODE LOOP
|
|
488
|
-
|
|
489
|
-
```
|
|
490
|
-
[8.1] WHILE tokensGenerated < maxTokens AND NOT stopped:
|
|
491
|
-
│
|
|
492
|
-
├─[8.1.1] _embed([lastToken])
|
|
493
|
-
│ └─ Same as prefill but numTokens=1
|
|
494
|
-
│
|
|
495
|
-
├─[8.1.2] FOR l = 0 to numLayers-1:
|
|
496
|
-
│ │
|
|
497
|
-
│ └─ _processLayerGPU(l, hiddenStates, numTokens=1, isPrefill=false)
|
|
498
|
-
│ │
|
|
499
|
-
│ └─ Attention uses DECODE kernel variant:
|
|
500
|
-
│ ├─ DEP: sharedMem → decode kernel (256 workgroup)
|
|
501
|
-
│ ├─ Reads full KV cache from previous positions
|
|
502
|
-
│ └─ Appends new K,V to cache
|
|
503
|
-
│
|
|
504
|
-
├─[8.1.3] _computeLogits(hiddenStates, 1)
|
|
505
|
-
│ └─ Output: logits[vocabSize] for position 0
|
|
506
|
-
│
|
|
507
|
-
├─[8.1.4] _sample(logits, opts)
|
|
508
|
-
│ └─ Output: nextToken
|
|
509
|
-
│
|
|
510
|
-
├─[8.1.5] Check stop conditions
|
|
511
|
-
│ ├─ stopTokenIds from modelConfig (model-specific)
|
|
512
|
-
│ └─ Check stopSequences in generated text
|
|
513
|
-
│
|
|
514
|
-
└─[8.1.6] currentSeqLen++
|
|
515
|
-
```
|
|
516
|
-
|
|
517
|
-
---
|
|
518
|
-
|
|
519
|
-
## External Dependencies Summary
|
|
520
|
-
|
|
521
|
-
| Dependency | Detection | Impact |
|
|
522
|
-
|------------|-----------|--------|
|
|
523
|
-
| **Memory64** | `getMemoryCapabilities().hasMemory64` | >4GB heap allocation |
|
|
524
|
-
| **Unified Memory** | `getMemoryCapabilities().isUnifiedMemory` | Dense model performance |
|
|
525
|
-
| **WebGPU** | `getDevice()` | GPU vs CPU path |
|
|
526
|
-
| **shader-f16** | `getKernelCapabilities().hasF16` | F16 matmul, F16 KV cache |
|
|
527
|
-
| **subgroups** | `getKernelCapabilities().hasSubgroups` | Fast dequantization |
|
|
528
|
-
| **Shared Memory** | `device.limits.maxComputeWorkgroupStorageSize` | Attention kernel variant |
|
|
529
|
-
| **OPFS** | Browser support | Model storage |
|
|
530
|
-
| **Native Bridge** | `storageContext.loadShard` | Bypass OPFS, mmap access |
|
|
531
|
-
|
|
532
|
-
---
|
|
533
|
-
|
|
534
|
-
## Model-Specific Forks Summary
|
|
535
|
-
|
|
536
|
-
| Step | Gemma 3 | LLaMA/Mistral | Mixtral |
|
|
537
|
-
|------|---------|---------------|---------|
|
|
538
|
-
| Tensor prefix | `language_model.model.layers` | `model.layers` | `model.layers` |
|
|
539
|
-
| Q/K norms | Yes | No | No |
|
|
540
|
-
| Embedding scale | `*= sqrt(hiddenSize)` | No | No |
|
|
541
|
-
| Norm architecture | Sandwich (4 norms) | Standard (2 norms) | Standard (2 norms) |
|
|
542
|
-
| Residual pattern | Norm BEFORE residual | Residual BEFORE norm | Residual BEFORE norm |
|
|
543
|
-
| FFN type | Dense | Dense | MoE (8 experts, top-2) |
|
|
544
|
-
| Sliding window | 1024 | None / 4096 (Mistral) | None |
|
|
545
|
-
| Stop tokens | From manifest | From manifest | From manifest |
|
|
546
|
-
| Chat template | Yes | No | No |
|
|
547
|
-
|
|
548
|
-
---
|
|
549
|
-
|
|
550
|
-
## Key Architecture Differences
|
|
551
|
-
|
|
552
|
-
### Sandwich Norm (Gemma 3) vs Standard (LLaMA/Mistral)
|
|
553
|
-
|
|
554
|
-
```
|
|
555
|
-
┌─────────────────────────────────────────────────────────────────────┐
|
|
556
|
-
│ STANDARD (LLaMA, Mistral, Mixtral) │
|
|
557
|
-
├─────────────────────────────────────────────────────────────────────┤
|
|
558
|
-
│ x = x + attn(input_norm(x)) ← residual wraps attention │
|
|
559
|
-
│ x = x + ffn(post_attn_norm(x)) ← residual wraps FFN │
|
|
560
|
-
└─────────────────────────────────────────────────────────────────────┘
|
|
561
|
-
|
|
562
|
-
┌─────────────────────────────────────────────────────────────────────┐
|
|
563
|
-
│ SANDWICH (Gemma 3) │
|
|
564
|
-
├─────────────────────────────────────────────────────────────────────┤
|
|
565
|
-
│ h = attn(input_norm(x)) │
|
|
566
|
-
│ h = post_attention_norm(h) ← norm ON attention output │
|
|
567
|
-
│ x = x + h ← then residual add │
|
|
568
|
-
│ h = ffn(pre_feedforward_norm(x)) │
|
|
569
|
-
│ h = post_feedforward_norm(h) ← norm ON FFN output │
|
|
570
|
-
│ x = x + h ← then residual add │
|
|
571
|
-
└─────────────────────────────────────────────────────────────────────┘
|
|
572
|
-
```
|
|
573
|
-
|
|
574
|
-
---
|
|
575
|
-
|
|
576
|
-
## Files Reference
|
|
577
|
-
|
|
578
|
-
| File | Purpose |
|
|
579
|
-
|------|---------|
|
|
580
|
-
| `inference/pipelines/text.js` | Main inference loop, layer processing |
|
|
581
|
-
| `loader/doppler-loader.js` | Weight loading, tensor naming |
|
|
582
|
-
| `gpu/kernel-selector.js` | Kernel dispatch based on capabilities |
|
|
583
|
-
| `memory/capability.js` | Memory64, unified memory detection |
|
|
584
|
-
| `inference/kv-cache.js` | KV cache management |
|
|
585
|
-
| `gpu/device.js` | WebGPU device, feature detection |
|
|
586
|
-
|
|
587
|
-
---
|
|
588
|
-
|
|
589
|
-
*Last updated: December 2025*
|
|
590
|
-
|
|
591
|
-
<!-- DOPPLER_KERNEL_OVERRIDES -->
|
|
592
|
-
## Kernel Overrides & Compatibility
|
|
593
|
-
See `docs/operations.md#kernel-overrides--compatibility` (canonical section).
|
|
@@ -1,10 +0,0 @@
|
|
|
1
|
-
export declare function resolveDeviceInfo(): Record<string, unknown> | null;
|
|
2
|
-
export declare function resolveKernelPathForModel(options?: Record<string, unknown>): Promise<{
|
|
3
|
-
modelId: string | null;
|
|
4
|
-
kernelPath: unknown;
|
|
5
|
-
source: string | null;
|
|
6
|
-
} | null>;
|
|
7
|
-
export declare function resolveLocalSourceRuntimePathFromModelUrl(
|
|
8
|
-
modelUrl?: string | null
|
|
9
|
-
): Promise<string | null>;
|
|
10
|
-
export declare function initializeSuiteModel(options?: Record<string, unknown>): Promise<Record<string, unknown>>;
|