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
|
@@ -7,6 +7,10 @@ import { acquireBuffer, releaseBuffer } from '../../../memory/buffer-pool.js';
|
|
|
7
7
|
|
|
8
8
|
const MAX_ADAM_ELEMENTS_PER_DISPATCH = 65535 * WORKGROUP_SIZES.DEFAULT;
|
|
9
9
|
|
|
10
|
+
function tensorElementCount(tensor) {
|
|
11
|
+
return tensor.shape.reduce((product, value) => product * value, 1);
|
|
12
|
+
}
|
|
13
|
+
|
|
10
14
|
async function runAdamChunked(device, pipeline, params, grads, moment1, moment2, options, inferredCount) {
|
|
11
15
|
if (params.dtype !== 'f32' || grads.dtype !== 'f32' || moment1.dtype !== 'f32' || moment2.dtype !== 'f32') {
|
|
12
16
|
throw new Error(
|
|
@@ -94,8 +98,7 @@ export async function runAdam(
|
|
|
94
98
|
const device = getDevice();
|
|
95
99
|
const { count, step = 1, lr, beta1, beta2, eps } = options;
|
|
96
100
|
|
|
97
|
-
const
|
|
98
|
-
const inferredCount = count ?? Math.floor(params.buffer.size / bytesPerElement);
|
|
101
|
+
const inferredCount = count ?? tensorElementCount(params);
|
|
99
102
|
const pipeline = await createPipeline('adam', 'default');
|
|
100
103
|
|
|
101
104
|
if (inferredCount > MAX_ADAM_ELEMENTS_PER_DISPATCH) {
|
|
@@ -166,8 +169,7 @@ export async function recordAdam(
|
|
|
166
169
|
const device = recorder.device;
|
|
167
170
|
const { count, step = 1, lr, beta1, beta2, eps } = options;
|
|
168
171
|
|
|
169
|
-
const
|
|
170
|
-
const inferredCount = count ?? Math.floor(params.buffer.size / bytesPerElement);
|
|
172
|
+
const inferredCount = count ?? tensorElementCount(params);
|
|
171
173
|
const pipeline = await createPipeline('adam', 'default');
|
|
172
174
|
|
|
173
175
|
const uniformBuffer = createUniformBufferWithView(
|
|
@@ -0,0 +1,24 @@
|
|
|
1
|
+
import type { CommandRecorder } from '../../command-recorder.js';
|
|
2
|
+
import type { Tensor } from '../../tensor.js';
|
|
3
|
+
|
|
4
|
+
export interface CausalConv1dSiluBackwardOptions {
|
|
5
|
+
numTokens: number;
|
|
6
|
+
channels: number;
|
|
7
|
+
kernelSize: number;
|
|
8
|
+
outputBuffer?: GPUBuffer | null;
|
|
9
|
+
}
|
|
10
|
+
|
|
11
|
+
export declare function runCausalConv1dSiluBackward(
|
|
12
|
+
input: Tensor,
|
|
13
|
+
weight: Tensor,
|
|
14
|
+
gradOutput: Tensor,
|
|
15
|
+
options: CausalConv1dSiluBackwardOptions
|
|
16
|
+
): Promise<Tensor>;
|
|
17
|
+
|
|
18
|
+
export declare function recordCausalConv1dSiluBackward(
|
|
19
|
+
recorder: CommandRecorder,
|
|
20
|
+
input: Tensor,
|
|
21
|
+
weight: Tensor,
|
|
22
|
+
gradOutput: Tensor,
|
|
23
|
+
options: CausalConv1dSiluBackwardOptions
|
|
24
|
+
): Promise<Tensor>;
|
|
@@ -0,0 +1,74 @@
|
|
|
1
|
+
import { getDevice } from '../../device.js';
|
|
2
|
+
import { acquireBuffer } from '../../../memory/buffer-pool.js';
|
|
3
|
+
import { createTensor } from '../../tensor.js';
|
|
4
|
+
import { WORKGROUP_SIZES } from '../constants.js';
|
|
5
|
+
import { dispatch, recordDispatch } from '../dispatch.js';
|
|
6
|
+
import { createPipeline, createUniformBufferWithView } from '../utils.js';
|
|
7
|
+
import { releaseUniformBuffer } from '../../uniform-cache.js';
|
|
8
|
+
|
|
9
|
+
function validate(input, weight, gradOutput, options) {
|
|
10
|
+
const numTokens = Math.floor(Number(options?.numTokens));
|
|
11
|
+
const channels = Math.floor(Number(options?.channels));
|
|
12
|
+
const kernelSize = Math.floor(Number(options?.kernelSize));
|
|
13
|
+
if (numTokens < 1 || channels < 1 || kernelSize < 1) {
|
|
14
|
+
throw new Error('causal conv1d SiLU backward requires numTokens, channels, and kernelSize.');
|
|
15
|
+
}
|
|
16
|
+
for (const [label, tensor] of [['input', input], ['weight', weight], ['gradOutput', gradOutput]]) {
|
|
17
|
+
if (tensor?.dtype !== 'f32') {
|
|
18
|
+
throw new Error(`causal conv1d SiLU backward requires f32 ${label}.`);
|
|
19
|
+
}
|
|
20
|
+
}
|
|
21
|
+
return { numTokens, channels, kernelSize };
|
|
22
|
+
}
|
|
23
|
+
|
|
24
|
+
async function execute(recorder, input, weight, gradOutput, options = {}) {
|
|
25
|
+
const dims = validate(input, weight, gradOutput, options);
|
|
26
|
+
const device = recorder?.device || getDevice();
|
|
27
|
+
if (!device) throw new Error('causal conv1d SiLU backward requires an active GPU device.');
|
|
28
|
+
const outputBuffer = options.outputBuffer
|
|
29
|
+
|| acquireBuffer(dims.numTokens * dims.channels * 4, undefined, 'causal_conv1d_silu_backward_output');
|
|
30
|
+
const pipeline = await createPipeline('causal_conv1d_silu_backward', 'default');
|
|
31
|
+
const uniformBuffer = createUniformBufferWithView(
|
|
32
|
+
'causal_conv1d_silu_backward_uniforms',
|
|
33
|
+
16,
|
|
34
|
+
(view) => {
|
|
35
|
+
view.setUint32(0, dims.numTokens, true);
|
|
36
|
+
view.setUint32(4, dims.channels, true);
|
|
37
|
+
view.setUint32(8, dims.kernelSize, true);
|
|
38
|
+
},
|
|
39
|
+
recorder || null,
|
|
40
|
+
device
|
|
41
|
+
);
|
|
42
|
+
const bindGroup = device.createBindGroup({
|
|
43
|
+
label: 'causal_conv1d_silu_backward_bind_group',
|
|
44
|
+
layout: pipeline.getBindGroupLayout(0),
|
|
45
|
+
entries: [
|
|
46
|
+
{ binding: 0, resource: { buffer: uniformBuffer } },
|
|
47
|
+
{ binding: 1, resource: { buffer: input.buffer } },
|
|
48
|
+
{ binding: 2, resource: { buffer: weight.buffer } },
|
|
49
|
+
{ binding: 3, resource: { buffer: gradOutput.buffer } },
|
|
50
|
+
{ binding: 4, resource: { buffer: outputBuffer } },
|
|
51
|
+
],
|
|
52
|
+
});
|
|
53
|
+
const workgroups = Math.ceil((dims.numTokens * dims.channels) / WORKGROUP_SIZES.DEFAULT);
|
|
54
|
+
if (recorder) {
|
|
55
|
+
recordDispatch(recorder, pipeline, bindGroup, workgroups, 'causal_conv1d_silu_backward');
|
|
56
|
+
} else {
|
|
57
|
+
dispatch(device, pipeline, bindGroup, workgroups, 'causal_conv1d_silu_backward');
|
|
58
|
+
releaseUniformBuffer(uniformBuffer);
|
|
59
|
+
}
|
|
60
|
+
return createTensor(
|
|
61
|
+
outputBuffer,
|
|
62
|
+
'f32',
|
|
63
|
+
[dims.numTokens, dims.channels],
|
|
64
|
+
'causal_conv1d_silu_backward_output'
|
|
65
|
+
);
|
|
66
|
+
}
|
|
67
|
+
|
|
68
|
+
export function runCausalConv1dSiluBackward(input, weight, gradOutput, options = {}) {
|
|
69
|
+
return execute(null, input, weight, gradOutput, options);
|
|
70
|
+
}
|
|
71
|
+
|
|
72
|
+
export function recordCausalConv1dSiluBackward(recorder, input, weight, gradOutput, options = {}) {
|
|
73
|
+
return execute(recorder, input, weight, gradOutput, options);
|
|
74
|
+
}
|
|
@@ -0,0 +1,59 @@
|
|
|
1
|
+
override WORKGROUP_SIZE: u32 = 256u;
|
|
2
|
+
|
|
3
|
+
struct Uniforms {
|
|
4
|
+
num_tokens: u32,
|
|
5
|
+
channels: u32,
|
|
6
|
+
kernel_size: u32,
|
|
7
|
+
_pad0: u32,
|
|
8
|
+
}
|
|
9
|
+
|
|
10
|
+
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
11
|
+
@group(0) @binding(1) var<storage, read> input: array<f32>;
|
|
12
|
+
@group(0) @binding(2) var<storage, read> weight: array<f32>;
|
|
13
|
+
@group(0) @binding(3) var<storage, read> grad_output: array<f32>;
|
|
14
|
+
@group(0) @binding(4) var<storage, read_write> grad_input: array<f32>;
|
|
15
|
+
|
|
16
|
+
fn sigmoid(x: f32) -> f32 {
|
|
17
|
+
if (x >= 0.0) {
|
|
18
|
+
let z = exp(-x);
|
|
19
|
+
return 1.0 / (1.0 + z);
|
|
20
|
+
}
|
|
21
|
+
let z = exp(x);
|
|
22
|
+
return z / (1.0 + z);
|
|
23
|
+
}
|
|
24
|
+
|
|
25
|
+
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
26
|
+
fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
|
|
27
|
+
let element = gid.x;
|
|
28
|
+
let element_count = u.num_tokens * u.channels;
|
|
29
|
+
if (element >= element_count) {
|
|
30
|
+
return;
|
|
31
|
+
}
|
|
32
|
+
|
|
33
|
+
let token = element / u.channels;
|
|
34
|
+
let channel = element % u.channels;
|
|
35
|
+
let last_output_token = min(u.num_tokens - 1u, token + u.kernel_size - 1u);
|
|
36
|
+
var gradient: f32 = 0.0;
|
|
37
|
+
|
|
38
|
+
for (var output_token = token; output_token <= last_output_token; output_token = output_token + 1u) {
|
|
39
|
+
let kernel_index = token + u.kernel_size - 1u - output_token;
|
|
40
|
+
var raw: f32 = 0.0;
|
|
41
|
+
for (var kernel: u32 = 0u; kernel < u.kernel_size; kernel = kernel + 1u) {
|
|
42
|
+
let padded_source = i32(output_token) + i32(kernel) - i32(u.kernel_size) + 1;
|
|
43
|
+
if (padded_source < 0) {
|
|
44
|
+
continue;
|
|
45
|
+
}
|
|
46
|
+
let source = u32(padded_source) * u.channels + channel;
|
|
47
|
+
raw = raw + input[source] * weight[channel * u.kernel_size + kernel];
|
|
48
|
+
}
|
|
49
|
+
let probability = sigmoid(raw);
|
|
50
|
+
let silu_derivative = probability * (1.0 + raw * (1.0 - probability));
|
|
51
|
+
let output_offset = output_token * u.channels + channel;
|
|
52
|
+
gradient = gradient
|
|
53
|
+
+ grad_output[output_offset]
|
|
54
|
+
* silu_derivative
|
|
55
|
+
* weight[channel * u.kernel_size + kernel_index];
|
|
56
|
+
}
|
|
57
|
+
|
|
58
|
+
grad_input[element] = gradient;
|
|
59
|
+
}
|
|
@@ -30,6 +30,10 @@ fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
|
|
|
30
30
|
let token_idx = idx / u.vocab_size;
|
|
31
31
|
let class_idx = idx % u.vocab_size;
|
|
32
32
|
let target_idx = targets[token_idx];
|
|
33
|
+
if (target_idx >= u.vocab_size) {
|
|
34
|
+
output[idx] = 0.0;
|
|
35
|
+
return;
|
|
36
|
+
}
|
|
33
37
|
var grad = softmax[idx];
|
|
34
38
|
if (class_idx == target_idx) {
|
|
35
39
|
grad = grad - 1.0;
|
|
@@ -0,0 +1,39 @@
|
|
|
1
|
+
import type { Tensor } from '../../tensor.js';
|
|
2
|
+
|
|
3
|
+
export interface GatedDeltaRecurrentBackwardInputs {
|
|
4
|
+
query: Tensor;
|
|
5
|
+
key: Tensor;
|
|
6
|
+
value: Tensor;
|
|
7
|
+
logDecay: Tensor;
|
|
8
|
+
beta: Tensor;
|
|
9
|
+
stateHistory: Tensor;
|
|
10
|
+
gradOutput: Tensor;
|
|
11
|
+
}
|
|
12
|
+
|
|
13
|
+
export interface GatedDeltaRecurrentBackwardOptions {
|
|
14
|
+
numTokens: number;
|
|
15
|
+
numHeads: number;
|
|
16
|
+
keyDim: number;
|
|
17
|
+
valueDim: number;
|
|
18
|
+
queryScale: number;
|
|
19
|
+
gradQueryBuffer?: GPUBuffer | null;
|
|
20
|
+
gradKeyBuffer?: GPUBuffer | null;
|
|
21
|
+
gradValueBuffer?: GPUBuffer | null;
|
|
22
|
+
gradLogDecayBuffer?: GPUBuffer | null;
|
|
23
|
+
gradBetaBuffer?: GPUBuffer | null;
|
|
24
|
+
gradStateBuffer?: GPUBuffer | null;
|
|
25
|
+
}
|
|
26
|
+
|
|
27
|
+
export interface GatedDeltaRecurrentBackwardResult {
|
|
28
|
+
query: Tensor;
|
|
29
|
+
key: Tensor;
|
|
30
|
+
value: Tensor;
|
|
31
|
+
logDecay: Tensor;
|
|
32
|
+
beta: Tensor;
|
|
33
|
+
initialState: Tensor;
|
|
34
|
+
}
|
|
35
|
+
|
|
36
|
+
export declare function runGatedDeltaRecurrentBackward(
|
|
37
|
+
inputs: GatedDeltaRecurrentBackwardInputs,
|
|
38
|
+
options: GatedDeltaRecurrentBackwardOptions
|
|
39
|
+
): Promise<GatedDeltaRecurrentBackwardResult>;
|
|
@@ -0,0 +1,124 @@
|
|
|
1
|
+
import { getDevice } from '../../device.js';
|
|
2
|
+
import { acquireBuffer, releaseBuffer } from '../../../memory/buffer-pool.js';
|
|
3
|
+
import { createTensor } from '../../tensor.js';
|
|
4
|
+
import { dispatch } from '../dispatch.js';
|
|
5
|
+
import { createPipeline, createUniformBufferWithView } from '../utils.js';
|
|
6
|
+
import { releaseUniformBuffer } from '../../uniform-cache.js';
|
|
7
|
+
|
|
8
|
+
function validate(inputs, options) {
|
|
9
|
+
const numTokens = Math.floor(Number(options?.numTokens));
|
|
10
|
+
const numHeads = Math.floor(Number(options?.numHeads));
|
|
11
|
+
const keyDim = Math.floor(Number(options?.keyDim));
|
|
12
|
+
const valueDim = Math.floor(Number(options?.valueDim));
|
|
13
|
+
const queryScale = Number(options?.queryScale);
|
|
14
|
+
if (numTokens < 1 || numHeads < 1 || keyDim < 1 || valueDim < 1 || valueDim > 128) {
|
|
15
|
+
throw new Error('gated-delta recurrent backward requires positive dimensions and valueDim <= 128.');
|
|
16
|
+
}
|
|
17
|
+
if (!Number.isFinite(queryScale)) {
|
|
18
|
+
throw new Error('gated-delta recurrent backward requires finite queryScale.');
|
|
19
|
+
}
|
|
20
|
+
for (const [label, tensor] of Object.entries(inputs)) {
|
|
21
|
+
if (tensor?.dtype !== 'f32') {
|
|
22
|
+
throw new Error(`gated-delta recurrent backward requires f32 ${label}.`);
|
|
23
|
+
}
|
|
24
|
+
}
|
|
25
|
+
return { numTokens, numHeads, keyDim, valueDim, queryScale };
|
|
26
|
+
}
|
|
27
|
+
|
|
28
|
+
function allocate(bytes, label, provided) {
|
|
29
|
+
return provided || acquireBuffer(bytes, undefined, label);
|
|
30
|
+
}
|
|
31
|
+
|
|
32
|
+
export async function runGatedDeltaRecurrentBackward(inputs, options = {}) {
|
|
33
|
+
const dims = validate(inputs, options);
|
|
34
|
+
const device = getDevice();
|
|
35
|
+
if (!device) throw new Error('gated-delta recurrent backward requires an active GPU device.');
|
|
36
|
+
const queryBytes = dims.numTokens * dims.numHeads * dims.keyDim * 4;
|
|
37
|
+
const valueBytes = dims.numTokens * dims.numHeads * dims.valueDim * 4;
|
|
38
|
+
const scalarBytes = dims.numTokens * dims.numHeads * 4;
|
|
39
|
+
const stateBytes = dims.numHeads * dims.keyDim * dims.valueDim * 4;
|
|
40
|
+
const packedQueryKeyBuffer = acquireBuffer(queryBytes * 2, undefined, 'gated_delta_packed_query_key');
|
|
41
|
+
const packedDecayBetaBuffer = acquireBuffer(scalarBytes * 2, undefined, 'gated_delta_packed_decay_beta');
|
|
42
|
+
const packedGradQueryKeyBuffer = acquireBuffer(queryBytes * 2, undefined, 'gated_delta_packed_grad_query_key');
|
|
43
|
+
const packedGradDecayBetaBuffer = acquireBuffer(scalarBytes * 2, undefined, 'gated_delta_packed_grad_decay_beta');
|
|
44
|
+
const gradQueryBuffer = allocate(queryBytes, 'gated_delta_grad_query', options.gradQueryBuffer);
|
|
45
|
+
const gradKeyBuffer = allocate(queryBytes, 'gated_delta_grad_key', options.gradKeyBuffer);
|
|
46
|
+
const gradValueBuffer = allocate(valueBytes, 'gated_delta_grad_value', options.gradValueBuffer);
|
|
47
|
+
const gradLogDecayBuffer = allocate(scalarBytes, 'gated_delta_grad_log_decay', options.gradLogDecayBuffer);
|
|
48
|
+
const gradBetaBuffer = allocate(scalarBytes, 'gated_delta_grad_beta', options.gradBetaBuffer);
|
|
49
|
+
const gradStateBuffer = allocate(stateBytes, 'gated_delta_grad_initial_state', options.gradStateBuffer);
|
|
50
|
+
let uniformBuffer = null;
|
|
51
|
+
let completed = false;
|
|
52
|
+
try {
|
|
53
|
+
const packEncoder = device.createCommandEncoder({ label: 'gated_delta_backward_pack_inputs' });
|
|
54
|
+
packEncoder.copyBufferToBuffer(inputs.query.buffer, 0, packedQueryKeyBuffer, 0, queryBytes);
|
|
55
|
+
packEncoder.copyBufferToBuffer(inputs.key.buffer, 0, packedQueryKeyBuffer, queryBytes, queryBytes);
|
|
56
|
+
packEncoder.copyBufferToBuffer(inputs.logDecay.buffer, 0, packedDecayBetaBuffer, 0, scalarBytes);
|
|
57
|
+
packEncoder.copyBufferToBuffer(inputs.beta.buffer, 0, packedDecayBetaBuffer, scalarBytes, scalarBytes);
|
|
58
|
+
device.queue.submit([packEncoder.finish()]);
|
|
59
|
+
device.queue.writeBuffer(gradStateBuffer, 0, new Uint8Array(stateBytes));
|
|
60
|
+
|
|
61
|
+
const pipeline = await createPipeline('gated_delta_recurrent_backward', 'default');
|
|
62
|
+
uniformBuffer = createUniformBufferWithView(
|
|
63
|
+
'gated_delta_recurrent_backward_uniforms',
|
|
64
|
+
32,
|
|
65
|
+
(view) => {
|
|
66
|
+
view.setUint32(0, dims.numTokens, true);
|
|
67
|
+
view.setUint32(4, dims.numHeads, true);
|
|
68
|
+
view.setUint32(8, dims.keyDim, true);
|
|
69
|
+
view.setUint32(12, dims.valueDim, true);
|
|
70
|
+
view.setFloat32(16, dims.queryScale, true);
|
|
71
|
+
},
|
|
72
|
+
null,
|
|
73
|
+
device
|
|
74
|
+
);
|
|
75
|
+
const bindGroup = device.createBindGroup({
|
|
76
|
+
label: 'gated_delta_recurrent_backward_bind_group',
|
|
77
|
+
layout: pipeline.getBindGroupLayout(0),
|
|
78
|
+
entries: [
|
|
79
|
+
{ binding: 0, resource: { buffer: uniformBuffer } },
|
|
80
|
+
{ binding: 1, resource: { buffer: packedQueryKeyBuffer } },
|
|
81
|
+
{ binding: 2, resource: { buffer: inputs.value.buffer } },
|
|
82
|
+
{ binding: 3, resource: { buffer: packedDecayBetaBuffer } },
|
|
83
|
+
{ binding: 4, resource: { buffer: inputs.stateHistory.buffer } },
|
|
84
|
+
{ binding: 5, resource: { buffer: inputs.gradOutput.buffer } },
|
|
85
|
+
{ binding: 6, resource: { buffer: packedGradQueryKeyBuffer } },
|
|
86
|
+
{ binding: 7, resource: { buffer: gradValueBuffer } },
|
|
87
|
+
{ binding: 8, resource: { buffer: packedGradDecayBetaBuffer } },
|
|
88
|
+
{ binding: 9, resource: { buffer: gradStateBuffer } },
|
|
89
|
+
],
|
|
90
|
+
});
|
|
91
|
+
dispatch(device, pipeline, bindGroup, dims.numHeads, 'gated_delta_recurrent_backward');
|
|
92
|
+
const unpackEncoder = device.createCommandEncoder({ label: 'gated_delta_backward_unpack_outputs' });
|
|
93
|
+
unpackEncoder.copyBufferToBuffer(packedGradQueryKeyBuffer, 0, gradQueryBuffer, 0, queryBytes);
|
|
94
|
+
unpackEncoder.copyBufferToBuffer(packedGradQueryKeyBuffer, queryBytes, gradKeyBuffer, 0, queryBytes);
|
|
95
|
+
unpackEncoder.copyBufferToBuffer(packedGradDecayBetaBuffer, 0, gradLogDecayBuffer, 0, scalarBytes);
|
|
96
|
+
unpackEncoder.copyBufferToBuffer(packedGradDecayBetaBuffer, scalarBytes, gradBetaBuffer, 0, scalarBytes);
|
|
97
|
+
device.queue.submit([unpackEncoder.finish()]);
|
|
98
|
+
await device.queue.onSubmittedWorkDone();
|
|
99
|
+
completed = true;
|
|
100
|
+
return {
|
|
101
|
+
query: createTensor(gradQueryBuffer, 'f32', [dims.numTokens, dims.numHeads, dims.keyDim], 'gated_delta_grad_query'),
|
|
102
|
+
key: createTensor(gradKeyBuffer, 'f32', [dims.numTokens, dims.numHeads, dims.keyDim], 'gated_delta_grad_key'),
|
|
103
|
+
value: createTensor(gradValueBuffer, 'f32', [dims.numTokens, dims.numHeads, dims.valueDim], 'gated_delta_grad_value'),
|
|
104
|
+
logDecay: createTensor(gradLogDecayBuffer, 'f32', [dims.numTokens, dims.numHeads], 'gated_delta_grad_log_decay'),
|
|
105
|
+
beta: createTensor(gradBetaBuffer, 'f32', [dims.numTokens, dims.numHeads], 'gated_delta_grad_beta'),
|
|
106
|
+
initialState: createTensor(gradStateBuffer, 'f32', [dims.numHeads, dims.keyDim, dims.valueDim], 'gated_delta_grad_initial_state'),
|
|
107
|
+
};
|
|
108
|
+
} finally {
|
|
109
|
+
await device.queue.onSubmittedWorkDone().catch(() => {});
|
|
110
|
+
if (uniformBuffer) releaseUniformBuffer(uniformBuffer);
|
|
111
|
+
releaseBuffer(packedQueryKeyBuffer);
|
|
112
|
+
releaseBuffer(packedDecayBetaBuffer);
|
|
113
|
+
releaseBuffer(packedGradQueryKeyBuffer);
|
|
114
|
+
releaseBuffer(packedGradDecayBetaBuffer);
|
|
115
|
+
if (!completed) {
|
|
116
|
+
if (!options.gradQueryBuffer) releaseBuffer(gradQueryBuffer);
|
|
117
|
+
if (!options.gradKeyBuffer) releaseBuffer(gradKeyBuffer);
|
|
118
|
+
if (!options.gradValueBuffer) releaseBuffer(gradValueBuffer);
|
|
119
|
+
if (!options.gradLogDecayBuffer) releaseBuffer(gradLogDecayBuffer);
|
|
120
|
+
if (!options.gradBetaBuffer) releaseBuffer(gradBetaBuffer);
|
|
121
|
+
if (!options.gradStateBuffer) releaseBuffer(gradStateBuffer);
|
|
122
|
+
}
|
|
123
|
+
}
|
|
124
|
+
}
|
|
@@ -0,0 +1,177 @@
|
|
|
1
|
+
override WORKGROUP_SIZE: u32 = 128u;
|
|
2
|
+
const MAX_WORKGROUP_SIZE: u32 = 128u;
|
|
3
|
+
|
|
4
|
+
struct Uniforms {
|
|
5
|
+
num_tokens: u32,
|
|
6
|
+
num_heads: u32,
|
|
7
|
+
key_dim: u32,
|
|
8
|
+
value_dim: u32,
|
|
9
|
+
query_scale: f32,
|
|
10
|
+
_pad0: u32,
|
|
11
|
+
_pad1: u32,
|
|
12
|
+
_pad2: u32,
|
|
13
|
+
}
|
|
14
|
+
|
|
15
|
+
@group(0) @binding(0) var<uniform> u: Uniforms;
|
|
16
|
+
@group(0) @binding(1) var<storage, read> query_key: array<f32>;
|
|
17
|
+
@group(0) @binding(2) var<storage, read> value: array<f32>;
|
|
18
|
+
@group(0) @binding(3) var<storage, read> decay_beta: array<f32>;
|
|
19
|
+
@group(0) @binding(4) var<storage, read> state_history: array<f32>;
|
|
20
|
+
@group(0) @binding(5) var<storage, read> grad_output: array<f32>;
|
|
21
|
+
@group(0) @binding(6) var<storage, read_write> grad_query_key: array<f32>;
|
|
22
|
+
@group(0) @binding(7) var<storage, read_write> grad_value: array<f32>;
|
|
23
|
+
@group(0) @binding(8) var<storage, read_write> grad_decay_beta: array<f32>;
|
|
24
|
+
@group(0) @binding(9) var<storage, read_write> grad_state: array<f32>;
|
|
25
|
+
|
|
26
|
+
var<workgroup> shared_reduce: array<f32, MAX_WORKGROUP_SIZE>;
|
|
27
|
+
|
|
28
|
+
fn vector_index(token: u32, head: u32, dim: u32, width: u32) -> u32 {
|
|
29
|
+
return ((token * u.num_heads + head) * width) + dim;
|
|
30
|
+
}
|
|
31
|
+
|
|
32
|
+
fn state_history_index(token: u32, head: u32, key_index: u32, value_index: u32) -> u32 {
|
|
33
|
+
return ((((token * u.num_heads + head) * u.key_dim) + key_index) * u.value_dim) + value_index;
|
|
34
|
+
}
|
|
35
|
+
|
|
36
|
+
fn state_gradient_index(head: u32, key_index: u32, value_index: u32) -> u32 {
|
|
37
|
+
return (((head * u.key_dim) + key_index) * u.value_dim) + value_index;
|
|
38
|
+
}
|
|
39
|
+
|
|
40
|
+
@compute @workgroup_size(WORKGROUP_SIZE, 1, 1)
|
|
41
|
+
fn main(
|
|
42
|
+
@builtin(workgroup_id) wid: vec3<u32>,
|
|
43
|
+
@builtin(local_invocation_id) lid: vec3<u32>
|
|
44
|
+
) {
|
|
45
|
+
let head = wid.x;
|
|
46
|
+
let lane = lid.x;
|
|
47
|
+
if (head >= u.num_heads) {
|
|
48
|
+
return;
|
|
49
|
+
}
|
|
50
|
+
let is_active = lane < u.value_dim;
|
|
51
|
+
let query_elements = u.num_tokens * u.num_heads * u.key_dim;
|
|
52
|
+
let scalar_elements = u.num_tokens * u.num_heads;
|
|
53
|
+
|
|
54
|
+
for (var reverse_token: u32 = 0u; reverse_token < u.num_tokens; reverse_token = reverse_token + 1u) {
|
|
55
|
+
let token = u.num_tokens - 1u - reverse_token;
|
|
56
|
+
let scalar_index = token * u.num_heads + head;
|
|
57
|
+
let decay = exp(decay_beta[scalar_index]);
|
|
58
|
+
var memory: f32 = 0.0;
|
|
59
|
+
if (is_active) {
|
|
60
|
+
for (var key_index: u32 = 0u; key_index < u.key_dim; key_index = key_index + 1u) {
|
|
61
|
+
let previous_state = state_history[
|
|
62
|
+
state_history_index(token, head, key_index, lane)
|
|
63
|
+
];
|
|
64
|
+
memory = memory
|
|
65
|
+
+ previous_state
|
|
66
|
+
* decay
|
|
67
|
+
* query_key[query_elements + vector_index(token, head, key_index, u.key_dim)];
|
|
68
|
+
}
|
|
69
|
+
}
|
|
70
|
+
let value_offset = vector_index(token, head, lane, u.value_dim);
|
|
71
|
+
let value_at_token = select(0.0, value[value_offset], is_active);
|
|
72
|
+
let delta = (value_at_token - memory) * decay_beta[scalar_elements + scalar_index];
|
|
73
|
+
let output_gradient = select(0.0, grad_output[value_offset], is_active);
|
|
74
|
+
|
|
75
|
+
for (var key_index: u32 = 0u; key_index < u.key_dim; key_index = key_index + 1u) {
|
|
76
|
+
let query_offset = vector_index(token, head, key_index, u.key_dim);
|
|
77
|
+
let next_state = select(
|
|
78
|
+
0.0,
|
|
79
|
+
state_history[state_history_index(token + 1u, head, key_index, lane)],
|
|
80
|
+
is_active
|
|
81
|
+
);
|
|
82
|
+
shared_reduce[lane] = output_gradient * next_state * u.query_scale;
|
|
83
|
+
workgroupBarrier();
|
|
84
|
+
for (var stride: u32 = WORKGROUP_SIZE / 2u; stride > 0u; stride = stride / 2u) {
|
|
85
|
+
if (lane < stride) {
|
|
86
|
+
shared_reduce[lane] = shared_reduce[lane] + shared_reduce[lane + stride];
|
|
87
|
+
}
|
|
88
|
+
workgroupBarrier();
|
|
89
|
+
}
|
|
90
|
+
if (lane == 0u) {
|
|
91
|
+
grad_query_key[query_offset] = shared_reduce[0];
|
|
92
|
+
}
|
|
93
|
+
if (is_active) {
|
|
94
|
+
let gradient_state_offset = state_gradient_index(head, key_index, lane);
|
|
95
|
+
grad_state[gradient_state_offset] = grad_state[gradient_state_offset]
|
|
96
|
+
+ output_gradient
|
|
97
|
+
* query_key[query_offset]
|
|
98
|
+
* u.query_scale;
|
|
99
|
+
}
|
|
100
|
+
workgroupBarrier();
|
|
101
|
+
}
|
|
102
|
+
|
|
103
|
+
var grad_delta: f32 = 0.0;
|
|
104
|
+
if (is_active) {
|
|
105
|
+
for (var key_index: u32 = 0u; key_index < u.key_dim; key_index = key_index + 1u) {
|
|
106
|
+
let gradient_state_offset = state_gradient_index(head, key_index, lane);
|
|
107
|
+
grad_delta = grad_delta
|
|
108
|
+
+ grad_state[gradient_state_offset]
|
|
109
|
+
* query_key[query_elements + vector_index(token, head, key_index, u.key_dim)];
|
|
110
|
+
}
|
|
111
|
+
grad_value[value_offset] = grad_delta * decay_beta[scalar_elements + scalar_index];
|
|
112
|
+
}
|
|
113
|
+
let grad_memory = -grad_delta * decay_beta[scalar_elements + scalar_index];
|
|
114
|
+
shared_reduce[lane] = select(0.0, grad_delta * (value_at_token - memory), is_active);
|
|
115
|
+
workgroupBarrier();
|
|
116
|
+
for (var stride: u32 = WORKGROUP_SIZE / 2u; stride > 0u; stride = stride / 2u) {
|
|
117
|
+
if (lane < stride) {
|
|
118
|
+
shared_reduce[lane] = shared_reduce[lane] + shared_reduce[lane + stride];
|
|
119
|
+
}
|
|
120
|
+
workgroupBarrier();
|
|
121
|
+
}
|
|
122
|
+
if (lane == 0u) {
|
|
123
|
+
grad_decay_beta[scalar_elements + scalar_index] = shared_reduce[0];
|
|
124
|
+
}
|
|
125
|
+
workgroupBarrier();
|
|
126
|
+
|
|
127
|
+
var local_grad_decay: f32 = 0.0;
|
|
128
|
+
for (var key_index: u32 = 0u; key_index < u.key_dim; key_index = key_index + 1u) {
|
|
129
|
+
let key_offset = vector_index(token, head, key_index, u.key_dim);
|
|
130
|
+
let gradient_state_offset = state_gradient_index(head, key_index, lane);
|
|
131
|
+
let previous_state = select(
|
|
132
|
+
0.0,
|
|
133
|
+
state_history[state_history_index(token, head, key_index, lane)],
|
|
134
|
+
is_active
|
|
135
|
+
);
|
|
136
|
+
let decayed_state = previous_state * decay;
|
|
137
|
+
let gradient_before_decay_update = select(0.0, grad_state[gradient_state_offset], is_active);
|
|
138
|
+
|
|
139
|
+
shared_reduce[lane] = select(
|
|
140
|
+
0.0,
|
|
141
|
+
(gradient_before_decay_update * delta) + (grad_memory * decayed_state),
|
|
142
|
+
is_active
|
|
143
|
+
);
|
|
144
|
+
workgroupBarrier();
|
|
145
|
+
for (var stride: u32 = WORKGROUP_SIZE / 2u; stride > 0u; stride = stride / 2u) {
|
|
146
|
+
if (lane < stride) {
|
|
147
|
+
shared_reduce[lane] = shared_reduce[lane] + shared_reduce[lane + stride];
|
|
148
|
+
}
|
|
149
|
+
workgroupBarrier();
|
|
150
|
+
}
|
|
151
|
+
if (lane == 0u) {
|
|
152
|
+
grad_query_key[query_elements + key_offset] = shared_reduce[0];
|
|
153
|
+
}
|
|
154
|
+
|
|
155
|
+
if (is_active) {
|
|
156
|
+
let gradient_decayed_state = gradient_before_decay_update
|
|
157
|
+
+ grad_memory * query_key[query_elements + key_offset];
|
|
158
|
+
local_grad_decay = local_grad_decay + gradient_decayed_state * decayed_state;
|
|
159
|
+
grad_state[gradient_state_offset] = gradient_decayed_state * decay;
|
|
160
|
+
}
|
|
161
|
+
workgroupBarrier();
|
|
162
|
+
}
|
|
163
|
+
|
|
164
|
+
shared_reduce[lane] = select(0.0, local_grad_decay, is_active);
|
|
165
|
+
workgroupBarrier();
|
|
166
|
+
for (var stride: u32 = WORKGROUP_SIZE / 2u; stride > 0u; stride = stride / 2u) {
|
|
167
|
+
if (lane < stride) {
|
|
168
|
+
shared_reduce[lane] = shared_reduce[lane] + shared_reduce[lane + stride];
|
|
169
|
+
}
|
|
170
|
+
workgroupBarrier();
|
|
171
|
+
}
|
|
172
|
+
if (lane == 0u) {
|
|
173
|
+
grad_decay_beta[scalar_index] = shared_reduce[0];
|
|
174
|
+
}
|
|
175
|
+
workgroupBarrier();
|
|
176
|
+
}
|
|
177
|
+
}
|
|
@@ -0,0 +1,32 @@
|
|
|
1
|
+
import type { CommandRecorder } from '../../command-recorder.js';
|
|
2
|
+
import type { Tensor } from '../../tensor.js';
|
|
3
|
+
|
|
4
|
+
export interface GatedRmsNormBackwardOptions {
|
|
5
|
+
rows: number;
|
|
6
|
+
width: number;
|
|
7
|
+
eps: number;
|
|
8
|
+
gradInputBuffer?: GPUBuffer | null;
|
|
9
|
+
gradGateBuffer?: GPUBuffer | null;
|
|
10
|
+
}
|
|
11
|
+
|
|
12
|
+
export interface GatedRmsNormBackwardResult {
|
|
13
|
+
gradInput: Tensor;
|
|
14
|
+
gradGate: Tensor;
|
|
15
|
+
}
|
|
16
|
+
|
|
17
|
+
export declare function runGatedRmsNormBackward(
|
|
18
|
+
input: Tensor,
|
|
19
|
+
gate: Tensor,
|
|
20
|
+
weight: Tensor,
|
|
21
|
+
gradOutput: Tensor,
|
|
22
|
+
options: GatedRmsNormBackwardOptions
|
|
23
|
+
): Promise<GatedRmsNormBackwardResult>;
|
|
24
|
+
|
|
25
|
+
export declare function recordGatedRmsNormBackward(
|
|
26
|
+
recorder: CommandRecorder,
|
|
27
|
+
input: Tensor,
|
|
28
|
+
gate: Tensor,
|
|
29
|
+
weight: Tensor,
|
|
30
|
+
gradOutput: Tensor,
|
|
31
|
+
options: GatedRmsNormBackwardOptions
|
|
32
|
+
): Promise<GatedRmsNormBackwardResult>;
|
|
@@ -0,0 +1,90 @@
|
|
|
1
|
+
import { getDevice } from '../../device.js';
|
|
2
|
+
import { acquireBuffer } from '../../../memory/buffer-pool.js';
|
|
3
|
+
import { createTensor } from '../../tensor.js';
|
|
4
|
+
import { dispatch, recordDispatch } from '../dispatch.js';
|
|
5
|
+
import { createPipeline, createUniformBufferWithView } from '../utils.js';
|
|
6
|
+
import { releaseUniformBuffer } from '../../uniform-cache.js';
|
|
7
|
+
|
|
8
|
+
function validate(input, gate, weight, gradOutput, options) {
|
|
9
|
+
const rows = Math.floor(Number(options?.rows));
|
|
10
|
+
const width = Math.floor(Number(options?.width));
|
|
11
|
+
const eps = Number(options?.eps);
|
|
12
|
+
if (rows < 1 || width < 1 || !Number.isFinite(eps) || eps <= 0) {
|
|
13
|
+
throw new Error('gated RMSNorm backward requires rows, width, and positive eps.');
|
|
14
|
+
}
|
|
15
|
+
for (const [label, tensor] of [
|
|
16
|
+
['input', input],
|
|
17
|
+
['gate', gate],
|
|
18
|
+
['weight', weight],
|
|
19
|
+
['gradOutput', gradOutput],
|
|
20
|
+
]) {
|
|
21
|
+
if (tensor?.dtype !== 'f32') {
|
|
22
|
+
throw new Error(`gated RMSNorm backward requires f32 ${label}.`);
|
|
23
|
+
}
|
|
24
|
+
}
|
|
25
|
+
return { rows, width, eps };
|
|
26
|
+
}
|
|
27
|
+
|
|
28
|
+
async function execute(recorder, input, gate, weight, gradOutput, options = {}) {
|
|
29
|
+
const dims = validate(input, gate, weight, gradOutput, options);
|
|
30
|
+
const device = recorder?.device || getDevice();
|
|
31
|
+
if (!device) throw new Error('gated RMSNorm backward requires an active GPU device.');
|
|
32
|
+
const outputBytes = dims.rows * dims.width * 4;
|
|
33
|
+
const gradInputBuffer = options.gradInputBuffer
|
|
34
|
+
|| acquireBuffer(outputBytes, undefined, 'gated_rmsnorm_backward_input');
|
|
35
|
+
const gradGateBuffer = options.gradGateBuffer
|
|
36
|
+
|| acquireBuffer(outputBytes, undefined, 'gated_rmsnorm_backward_gate');
|
|
37
|
+
const pipeline = await createPipeline('gated_rmsnorm_backward', 'default');
|
|
38
|
+
const uniformBuffer = createUniformBufferWithView(
|
|
39
|
+
'gated_rmsnorm_backward_uniforms',
|
|
40
|
+
16,
|
|
41
|
+
(view) => {
|
|
42
|
+
view.setUint32(0, dims.rows, true);
|
|
43
|
+
view.setUint32(4, dims.width, true);
|
|
44
|
+
view.setFloat32(8, dims.eps, true);
|
|
45
|
+
},
|
|
46
|
+
recorder || null,
|
|
47
|
+
device
|
|
48
|
+
);
|
|
49
|
+
const bindGroup = device.createBindGroup({
|
|
50
|
+
label: 'gated_rmsnorm_backward_bind_group',
|
|
51
|
+
layout: pipeline.getBindGroupLayout(0),
|
|
52
|
+
entries: [
|
|
53
|
+
{ binding: 0, resource: { buffer: uniformBuffer } },
|
|
54
|
+
{ binding: 1, resource: { buffer: input.buffer } },
|
|
55
|
+
{ binding: 2, resource: { buffer: gate.buffer } },
|
|
56
|
+
{ binding: 3, resource: { buffer: weight.buffer } },
|
|
57
|
+
{ binding: 4, resource: { buffer: gradOutput.buffer } },
|
|
58
|
+
{ binding: 5, resource: { buffer: gradInputBuffer } },
|
|
59
|
+
{ binding: 6, resource: { buffer: gradGateBuffer } },
|
|
60
|
+
],
|
|
61
|
+
});
|
|
62
|
+
if (recorder) {
|
|
63
|
+
recordDispatch(recorder, pipeline, bindGroup, dims.rows, 'gated_rmsnorm_backward');
|
|
64
|
+
} else {
|
|
65
|
+
dispatch(device, pipeline, bindGroup, dims.rows, 'gated_rmsnorm_backward');
|
|
66
|
+
releaseUniformBuffer(uniformBuffer);
|
|
67
|
+
}
|
|
68
|
+
return {
|
|
69
|
+
gradInput: createTensor(
|
|
70
|
+
gradInputBuffer,
|
|
71
|
+
'f32',
|
|
72
|
+
[dims.rows, dims.width],
|
|
73
|
+
'gated_rmsnorm_backward_input'
|
|
74
|
+
),
|
|
75
|
+
gradGate: createTensor(
|
|
76
|
+
gradGateBuffer,
|
|
77
|
+
'f32',
|
|
78
|
+
[dims.rows, dims.width],
|
|
79
|
+
'gated_rmsnorm_backward_gate'
|
|
80
|
+
),
|
|
81
|
+
};
|
|
82
|
+
}
|
|
83
|
+
|
|
84
|
+
export function runGatedRmsNormBackward(input, gate, weight, gradOutput, options = {}) {
|
|
85
|
+
return execute(null, input, gate, weight, gradOutput, options);
|
|
86
|
+
}
|
|
87
|
+
|
|
88
|
+
export function recordGatedRmsNormBackward(recorder, input, gate, weight, gradOutput, options = {}) {
|
|
89
|
+
return execute(recorder, input, gate, weight, gradOutput, options);
|
|
90
|
+
}
|